R-TF-028-006 AI Release Report

Table of contents

Purpose
Scope
Related Documents
Package Details
Clinical Models: Detailed Specifications
Non-Clinical Models: Detailed Specifications
Technical Integration Details
Integration Verification Package
Release Verification
Release Checklist
Technical Support

Purpose

This document serves as the official release report for the AI algorithm package integrated into the Legit.Health Plus medical device, version v1.1.0.0. It provides comprehensive technical specifications for all AI models, including input/output formats, preprocessing and post-processing requirements, and integration guidelines necessary for the deployment team to implement the algorithms in the production environment.

This release report is prepared in accordance with MDR 2017/745 Annex II requirements for technical documentation, specifically addressing the software as a medical device (SaMD) component containing AI/ML algorithms.

Scope

This document covers all 59 AI models included in the Legit.Health Plus device:

54 Clinical Models: Directly fulfill the device's intended purpose of providing quantitative data on clinical signs and interpretative distribution of ICD categories
5 Non-Clinical Models: Support proper device functioning through quality assurance, preprocessing, and validation functions

Document ID	Title	Relationship
GP-028	AI Development	Governing procedure for AI development and release activities
R-TF-028-001	AI Description	Model specifications, performance requirements, clinical objectives
R-TF-028-002	AI Development Plan	Development methodology and validation strategy
R-TF-028-005	AI Development Report	Training results, validation outcomes, bias analysis
R-TF-028-010	AI V&V Checks	Verification checklist for release deliverables
R-TF-028-011	AI Risk Assessment	Risk analysis and mitigation measures

Package Details

Overview

The algorithm package is composed of multiple deep learning models and configuration files, organized to support the device's clinical and non-clinical functions.

The complete package is available in the designated repository, version v1.1.0.0. It contains:

Clinical Models:

1 ICD Category Distribution model (.pt format) with configuration file (.json)
1 Binary Indicator mapping matrix (.csv) for post-processing derivation
10 Visual Sign Intensity Quantification models (.pt format) with configurations
24 Wound Characteristic Assessment models (.pt format) with configurations
5 Lesion Quantification models (object detection, .pt format)
12 Surface Area Quantification models (segmentation, .pt format)
2 Pattern Identification models (classification, .pt format)

Non-Clinical Models:

1 Domain Validation model (.pt format)
1 Dermatology Image Quality Assessment (DIQA) model (.pt format)
1 Skin Surface Segmentation model (.pt format)
1 Body Surface Segmentation model (.pt format)
1 Head Detection model (.pt format)

Model File Format

All deep learning models are provided in PyTorch native format (.pt, .pth, or .ckpt files), ensuring:

Full preservation of model architecture, weights, and computational graph
Reliable deployment and reproducibility
Version control and traceability
Optimized inference using PyTorch runtime (CPU and GPU)
Compliance with industry standards for medical device software

PyTorch Runtime Requirements:

PyTorch version: >=1.12
Python version: >=3.8
CUDA version (for GPU): >=11.6 (optional)

Configuration Files

Each model is accompanied by a JSON configuration file specifying:

Model path and version
Input dimensions and preprocessing parameters
Output format and post-processing requirements
Class labels (where applicable)
Threshold values for clinical decision support

Clinical Models: Detailed Specifications

ICD Category Distribution and Binary Indicators

Model Classification: 🔬 Clinical Model

Reference: R-TF-028-001 AI/ML Description - ICD Category Distribution and Binary Indicators section

ICD Category Distribution Algorithm

Purpose

The ICD Category Distribution algorithm analyzes a dermatological image (clinical or dermoscopic) to provide an interpretative distribution of probabilities across 346 validated ICD-11 categories. The algorithm outputs the top-5 most probable conditions with their corresponding ICD-11 codes and confidence scores.

The performance testing of this algorithm is detailed in R-TF-028-005 AI Development Report.

Input Specifications

Parameter	Specification
Type	Single dermatological image (clinical or dermoscopic)
Format	RGB, uint8; standard image file formats (JPEG, PNG)
Resolution	Minimum 224×224 pixels; recommended ≥512×512 pixels
Color Space	sRGB

Output Specifications

Parameter	Specification
Type	Probability distribution vector
Dimension	346 values (one per validated ICD-11 category)
Data Type	`float32`
Value Range	`[0.0, 1.0]` per element; sum of all values = `1.0`

Output JSON Structure:

{
  "icd_distribution": {
    "entropy": 0.45,
    "top_5_predictions": [
      {"icd_code": "EA90.0", "name": "Psoriasis vulgaris", "probability": 0.72},
      {"icd_code": "EA80.0", "name": "Atopic dermatitis", "probability": 0.12},
      {"icd_code": "EA85.0", "name": "Seborrhoeic dermatitis", "probability": 0.08},
      {"icd_code": "EK90.0", "name": "Tinea corporis", "probability": 0.04},
      {"icd_code": "EA81.0", "name": "Contact dermatitis", "probability": 0.02}
    ],
    "full_distribution": [0.72, 0.12, 0.08, ..., 0.000001]
  }
}

Preprocessing Requirements

The raw input image must be preprocessed before being passed to the PyTorch model:

Resize: Scale the image to 384×384 pixels using bilinear interpolation
Normalize: Apply image normalization using the training set mean pixel value and standard deviation.
Data Type Conversion: Convert to float32
Tensor Format: CHW (Channels, Height, Width) ordering

Post-processing Requirements

To achieve the validated performance, Test-Time Augmentation (TTA) must be applied:

Create augmented versions of the preprocessed input image:
- Horizontal flip
- Vertical flip
- 90°, 180°, 270° rotation
- Histogram equalization (CLAHE)
Run inference on the original image and all augmented versions (5 total inferences)
Average the probability vectors element-wise to produce the final distribution
Apply temperature scaling (calibration parameter provided in config) for probability calibration
Extract top-5 predictions by sorting probabilities in descending order
Compute the entropy of the probability distribution to provide additional explainability

Configuration File

File: icd_distribution_config.json

{
  "model_path": "models/icd_distribution_v1.1.0.pt",
  "model_version": "1.1.0.0",
  "architecture": "convnextv2_base",
  "head_hidden_features": 768,
  "head_activation": "gelu",
  "temperature_scaling": 1.1690781116485596,
  "num_classes": 346,
  "classes_file": "icd_categories_v1.1.0.0.csv",
  "input_size": 384,
  "normalization": {
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225]
  }
}

Performance Summary

Metric	Result	Success Criterion
Top-1 Accuracy	0.658 (95% CI: 0.654-0.663)	≥ 0.50
Top-3 Accuracy	0.821 (95% CI: 0.817-0.825)	≥ 0.60
Top-5 Accuracy	0.864 (95% CI: 0.861-0.868)	≥ 0.70

Expected Results

s3://legit-health-plus/test-images/condition-classifier/melanoma_1.jpg

{
  "icd_distribution": {
  "entropy": 0.39412604460588385,
  "top_5_predictions": [
    {
      "icd_code": "2C30",
      "name": "Cutaneous melanoma",
      "probability": 70.807356
    },
    {
      "icd_code": "2F20.1",
      "name": "Atypical melanocytic nevus",
      "probability": 0.308577
    },
    {
      "icd_code": "2F20.Z",
      "name": "Melanocytic nevus",
      "probability": 0.293256
    },
    {
      "icd_code": "2C32",
      "name": "Basal cell carcinoma",
      "probability": 0.204755
    },
    {
      "icd_code": "2F21.0",
      "name": "Seborrhoeic keratosis",
      "probability": 0.173751
    },
    {
      "icd_code": "EK90.0&XH36H6",
      "name": "Actinic keratosis",
      "probability": 0.131219
    },
    {
      "icd_code": "2C31",
      "name": "Squamous cell carcinoma",
      "probability": 0.120039
    },
    {
      "icd_code": "ED90.0",
      "name": "Rosacea",
      "probability": 0.112667
    },
    {
      "icd_code": "2C30.3",
      "name": "Acral lentiginous melanoma (primary)",
      "probability": 0.111564
    },
    {
      "icd_code": "ED80.Z",
      "name": "Acne",
      "probability": 0.107905
    }
  ],
  "full_distribution": [
    {
      "icd_code": "EE10.1",
      "name": "Abnormality of nail surface",
      "probability": 0.000837
    },
    {
      "icd_code": "ND56.0&XJ652",
      "name": "Abrasion",
      "probability": 0.00098
    },
    {
      "icd_code": "ED51.0",
      "name": "Acanthosis nigricans",
      "probability": 0.000811
    }, ...
    {
      "icd_code": "EF40",
      "name": "Vasculitis or capillaritis involving the skin",
      "probability": 0.000789
    },
    {
      "icd_code": "EG00",
      "name": "Vasodilatation of extremities",
      "probability": 0.000769
    },
    {
      "icd_code": "EF20.0",
      "name": "Venous lake",
      "probability": 0.000837
    },
    {
      "icd_code": "BD74.3",
      "name": "Venous leg ulcer",
      "probability": 0.000821
    },
    {
      "icd_code": "ED63.0",
      "name": "Vitiligo",
      "probability": 0.000917
    },
    {
      "icd_code": "2E67.11",
      "name": "Vulvar Paget disease",
      "probability": 0.000784
    },
    {
      "icd_code": "9A06.4",
      "name": "Xanthelasma of eyelid",
      "probability": 0.000868
    },
    {
      "icd_code": "LD27.1",
      "name": "Xeroderma pigmentosum",
      "probability": 0.000923
    },
    {
      "icd_code": "ED54",
      "name": "Xerosis cutis or asteatosis",
      "probability": 0.000785
    },
    {
      "icd_code": "EE11.1",
      "name": "Yellow nail syndrome",
      "probability": 0.000758
    },
    {
      "icd_code": "1E+091",
      "name": "Zoster",
      "probability": 0.000829
    }
  ]
}
}

s3://legit-health-plus/test-images/condition-classifier/melanoma_2.jpg

{
  "icd_distribution": {
    "entropy": 0.4414921339232481,
    "top_5_predictions": [
      {
        "icd_code": "2C30",
        "name": "Cutaneous melanoma",
        "probability": 66.2925
      },
      {
        "icd_code": "2C30.3",
        "name": "Acral lentiginous melanoma (primary)",
        "probability": 0.913159
      },
      {
        "icd_code": "2F20.1",
        "name": "Atypical melanocytic nevus",
        "probability": 0.688098
      },
      {
        "icd_code": "2F20.Z",
        "name": "Melanocytic nevus",
        "probability": 0.371657
      },
      {
        "icd_code": "2C32",
        "name": "Basal cell carcinoma",
        "probability": 0.283395
      },
      {
        "icd_code": "EA89",
        "name": "Eczematous dermatitis",
        "probability": 0.211095
      },
      {
        "icd_code": "EK90.0&XH36H6",
        "name": "Actinic keratosis",
        "probability": 0.143392
      },
      {
        "icd_code": "2C31",
        "name": "Squamous cell carcinoma",
        "probability": 0.142958
      },
      {
        "icd_code": "EB60",
        "name": "Lichen sclerosus",
        "probability": 0.140493
      },
      {
        "icd_code": "-",
        "name": "Non-specific finding",
        "probability": 0.128925
      }
    ],
    "full_distribution": [
      {
        "icd_code": "EE10.1",
        "name": "Abnormality of nail surface",
        "probability": 0.000849
      },
      {
        "icd_code": "ND56.0&XJ652",
        "name": "Abrasion",
        "probability": 0.000911
      },
      {
        "icd_code": "ED51.0",
        "name": "Acanthosis nigricans",
        "probability": 0.001079
      },
      {
        "icd_code": "LB63",
        "name": "Accessory nipple",
        "probability": 0.000948
      },
      {
        "icd_code": "ED80.Z",
        "name": "Acne",
        "probability": 0.001127
      },
      {
        "icd_code": "ED54",
        "name": "Xerosis cutis or asteatosis",
        "probability": 0.00086
      },
      {
        "icd_code": "EE11.1",
        "name": "Yellow nail syndrome",
        "probability": 0.000766
      },
      {
        "icd_code": "1E+091",
        "name": "Zoster",
        "probability": 0.001044
      }
    ]
  }
}

s3://legit-health-plus/test-images/condition-classifier/melanoma_3.jpg

{
  "icd_distribution": {
    "entropy": 0.4101957026013219,
    "top_5_predictions": [
      {
        "icd_code": "2C30",
        "name": "Cutaneous melanoma",
        "probability": 69.331157
      },
      {
        "icd_code": "2C30.3",
        "name": "Acral lentiginous melanoma (primary)",
        "probability": 0.416621
      },
      {
        "icd_code": "2F20.1",
        "name": "Atypical melanocytic nevus",
        "probability": 0.411692
      },
      {
        "icd_code": "2F20.Z",
        "name": "Melanocytic nevus",
        "probability": 0.248993
      },
      {
        "icd_code": "2C32",
        "name": "Basal cell carcinoma",
        "probability": 0.221238
      },
      {
        "icd_code": "EA89",
        "name": "Eczematous dermatitis",
        "probability": 0.142786
      },
      {
        "icd_code": "EK90.0&XH36H6",
        "name": "Actinic keratosis",
        "probability": 0.127115
      },
      {
        "icd_code": "2C31",
        "name": "Squamous cell carcinoma",
        "probability": 0.123789
      },
      {
        "icd_code": "LA90.1",
        "name": "Lymphatic malformation",
        "probability": 0.120748
      },
      {
        "icd_code": "2F21.0",
        "name": "Seborrhoeic keratosis",
        "probability": 0.118404
      }
    ],
    "full_distribution": [
      {
        "icd_code": "EE10.1",
        "name": "Abnormality of nail surface",
        "probability": 0.000784
      },
      {
        "icd_code": "ND56.0&XJ652",
        "name": "Abrasion",
        "probability": 0.000865
      },
      {
        "icd_code": "ED51.0",
        "name": "Acanthosis nigricans",
        "probability": 0.000993
      },
      {
        "icd_code": "9A06.4",
        "name": "Xanthelasma of eyelid",
        "probability": 0.000935
      },
      {
        "icd_code": "LD27.1",
        "name": "Xeroderma pigmentosum",
        "probability": 0.001066
      },
      {
        "icd_code": "ED54",
        "name": "Xerosis cutis or asteatosis",
        "probability": 0.000815
      },
      {
        "icd_code": "EE11.1",
        "name": "Yellow nail syndrome",
        "probability": 0.000745
      },
      {
        "icd_code": "1E+091",
        "name": "Zoster",
        "probability": 0.000954
      }
    ]
  }
}

Binary Indicators Algorithm

Purpose

Binary indicators are derived from the ICD-11 probability distribution as a deterministic post-processing step using a dermatologist-defined mapping matrix. Each indicator reflects the aggregated probability that a case belongs to clinically meaningful categories requiring differential triage or diagnostic attention.

Input Specifications

Parameter	Specification
Type	ICD probability distribution vector
Dimension	346 values (from ICD Category Distribution output)
Additional Input	Binary indicator mapping matrix (from CSV)

Output Specifications

Six Binary Indicators:

Indicator	Description	Clinical Significance
Malignant	Probability of confirmed malignancy	Melanoma, SCC, BCC detection
Pre-malignant	Probability of conditions with malignant potential	Actinic keratosis, Bowen's disease
Associated with Malignancy	Benign conditions mimicking malignancy	Atypical nevi, seborrheic keratoses
Pigmented Lesion	Probability of pigmented subgroup	Melanoma risk assessment
Urgent Referral	Requires evaluation within 48 hours	Suspected melanoma, ulcerated lesions
High-Priority Referral	Requires evaluation within 2 weeks	Suspected NMSC, premalignant lesions

Output JSON Structure:

{
  "binary_indicators": {
    "malignant": 0.03,
    "pre_malignant": 0.08,
    "associated_with_malignancy": 0.15,
    "pigmented_lesion": 0.92,
    "urgent_referral": 0.05,
    "high_priority_referral": 0.12
  }
}

Calculation Method

For each binary indicator $j$ , the integration code must perform:

\text{Binary Indicator}_j = \sum_{i=1}^{346} \big(p_i \times M_{ij}\big)

Where:

$p_i$ is the ICD-11 probability for category $i$
$M_{ij}$ is the binary mapping matrix coefficient (0 or 1) indicating whether category $i$ contributes to indicator $j$

Configuration File

File: icd_categories_v1.1.0.0.csv, with columns:

Class name (string)
ICD-11 code (string)
One column per binary indicator (int)

Performance Summary

Indicator	AUC	95% CI	Success Criterion
Malignant	0.918	0.914-0.922	≥ 0.80
Pre-malignant	0.878	0.872-0.884	≥ 0.80
Associated with Malignancy	0.863	0.855-0.870	≥ 0.80
Pigmented Lesion	0.959	0.957-0.962	≥ 0.80
Urgent Referral	0.900	0.889-0.911	≥ 0.80
High-Priority Referral	0.888	0.884-0.892	≥ 0.80

Expected Results

s3://legit-health-plus/test-images/condition-classifier/melanoma_1.jpg

{
  "binary_indicators": {
    "Malignant": 0.726005,
    "Pre-malignant": 0.034555,
    "Associated to malignancy": 0.768963,
    "Is a pigmented lesion": 0.759136,
    "Urgent referral": 0.049306,
    "High-priority referral": 0.797724
  }
}

s3://legit-health-plus/test-images/condition-classifier/melanoma_2.jpg

{
  "binary_indicators": {
    "Malignant": 0.691468,
    "Pre-malignant": 0.043102,
    "Associated to malignancy": 0.738586,
    "Is a pigmented lesion": 0.732039,
    "Urgent referral": 0.054326,
    "High-priority referral": 0.7752
  }
}

s3://legit-health-plus/test-images/condition-classifier/melanoma_3.jpg

{
  "binary_indicators": {
    "Malignant": 0.715312,
    "Pre-malignant": 0.037663,
    "Associated to malignancy": 0.759875,
    "Is a pigmented lesion": 0.749356,
    "Urgent referral": 0.051053,
    "High-priority referral": 0.791029
  }
}

Visual Sign Intensity Quantification Models

Model Classification: 🔬 Clinical Models

These models quantify the intensity of clinical signs on an ordinal scale (0-9), outputting probability distributions that are converted to continuous severity scores.

Common Specifications

Input Specifications (All Intensity Models)

Parameter	Specification
Type	Single dermatological image
Format	RGB, uint8 (JPEG, PNG)
Resolution	Resized to 272×272 pixels
Color Space	sRGB

Output Specifications (All Intensity Models)

Parameter	Specification
Probability Distribution	10 values (classes 0-9), sum = 1.0
Continuous Score	Weighted expected value: $\hat{y} = \sum_{i=0}^{9} i \cdot p_i$
Data Type	`float32`
Value Range	`[0.0, 9.0]`

Preprocessing (All Intensity Models)

Resize: Scale to 272×272 pixels using bilinear interpolation
Normalize: Apply ImageNet normalization:
- Mean: [0.485, 0.456, 0.406]
- Std: [0.229, 0.224, 0.225]
Data Type: Convert to float32
Tensor Format: CHW ordering

Post-processing (All Intensity Models)

Apply softmax activation to obtain probability distribution
Calculate continuous score: $\hat{y} = \sum_{i=0}^{9} i \cdot p_i$
Round to desired precision (typically 2 decimal places)

Architecture

All intensity models use EfficientNet-B2 with Feature Pyramid Network (FPN) backbone, adapted for 10-class ordinal classification.

Erythema Intensity Quantification

Purpose: Quantifies redness/inflammation intensity for conditions including psoriasis, atopic dermatitis, rosacea.

Expected output JSON:

{
  "probability_distribution": [
    [
      3.317265750979459e-8, 1.300240626278537e-7, 4.289904431686864e-9, 2.1801518812480936e-8,
      9.524950428385637e-7, 0.033597446978092194, 0.10959780961275101, 0.8562615513801575,
      0.0005419636145234108, 4.4181707181678576e-8
    ]
  ],
  "continuous_score": 6.8237450727682205
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/erythema-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Erythema Intensity Quantification

Configuration File: erythema_intensity_config.json

Desquamation Intensity Quantification

Purpose: Quantifies scaling/peeling intensity for psoriasis, seborrheic dermatitis, and other scaling conditions.

Expected output JSON:

{
  "probability_distribution": [
    [
      8.442028160970949e-7, 2.9521125100018253e-8, 3.969607860199176e-6, 0.0013594592455774546,
      0.5552752017974854, 4.094568430446088e-5, 0.0012707337737083435, 0.44140952825546265,
      0.0006375949014909565, 1.637903892515169e-6
    ]
  ],
  "continuous_score": 5.3279984828624904
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/desquamation-classifier/case-001/

Performance: Specified in R-TF-028-005 AI Development Report: Desquamation Intensity Quantification

Configuration File: desquamation_intensity_config.json

Induration Intensity Quantification

Purpose: Quantifies plaque thickness/hardening for psoriasis and other inflammatory dermatoses.

Expected output JSON:

{
  "probability_distribution": [
    [
      0.00338586769066751, 6.331619601951388e-7, 9.158335160464048e-5, 0.05288663133978844,
      0.006485740654170513, 0.00019502690702211112, 0.004077768884599209, 0.9296616911888123,
      0.0031048504170030355, 0.00011023339175153524
    ]
  ],
  "continuous_score": 6.7436911465273965
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/induration-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Induration Intensity Quantification

Configuration File: induration_intensity_config.json

Pustule Intensity Quantification

Purpose: Quantifies pustulation severity for pustular psoriasis, acne, AGEP.

Expected output JSON:

{
  "probability_distribution": [
    [
      0.1450415998697281, 0.006939078215509653, 0.7252118587493896, 0.003976427949965,
      0.00577736459672451, 0.001695547834970057, 0.007374573964625597, 0.07776667177677155,
      0.007189963944256306, 0.019026942551136017
    ]
  ],
  "continuous_score": 2.3182556178653613
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/pustule-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Pustule Intensity Quantification

Configuration File: pustule_intensity_config.json

Crusting Intensity Quantification

Purpose: Quantifies crust formation for atopic dermatitis, impetigo, wound healing assessment.

Expected output JSON:

{
  "probability_distribution": [
    [
      8.442028160970949e-7, 2.9521125100018253e-8, 3.969607860199176e-6, 0.0013594592455774546,
      0.5552752017974854, 4.094568430446088e-5, 0.0012707337737083435, 0.44140952825546265,
      0.0006375949014909565, 1.637903892515169e-6
    ]
  ],
  "continuous_score": 5.3279984828624904
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/crusting-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Crusting Intensity Quantification

Configuration File: crusting_intensity_config.json

Xerosis Intensity Quantification

Purpose: Quantifies dry skin severity for atopic dermatitis, ichthyosis, aging skin.

Expected output JSON:

{
  "probability_distribution": [
    [
      0.00048537287511862814, 0.04049737751483917, 0.019379710778594017, 0.001857674797065556,
      0.00037739728577435017, 6.432881491491571e-5, 0.9372596740722656, 5.054048597230576e-5,
      2.6538469683146104e-5, 1.4749366528121755e-6
    ]
  ],
  "continuous_score": 5.710798466703636
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/xerosis-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Xerosis Intensity Quantification

Configuration File: xerosis_intensity_config.json

Swelling Intensity Quantification

Purpose: Quantifies edema severity for atopic dermatitis, urticaria, angioedema.

Expected output JSON:

{
  "probability_distribution": [
    [
      0.000630750902928412, 0.0012454602401703596, 0.5647211670875549, 0.3996390700340271,
      0.003440126543864608, 0.024147767573595047, 0.0028311503119766712, 0.0027287527918815613,
      0.0005960240378044546, 1.975861050595995e-5
    ]
  ],
  "continuous_score": 2.5051385397728154
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/swelling-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Swelling Intensity Quantification

Configuration File: swelling_intensity_config.json

Oozing Intensity Quantification

Purpose: Quantifies exudate/discharge severity for acute dermatitis, infected eczema.

Expected output JSON:

{
  "probability_distribution": [
    [
      8.442028160970949e-7, 2.9521125100018253e-8, 3.969607860199176e-6, 0.0013594592455774546,
      0.5552752017974854, 4.094568430446088e-5, 0.0012707337737083435, 0.44140952825546265,
      0.0006375949014909565, 1.637903892515169e-6
    ]
  ],
  "continuous_score": 5.3279984828624904
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/oozing-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Oozing Intensity Quantification

Configuration File: oozing_intensity_config.json

Excoriation Intensity Quantification

Purpose: Quantifies scratch damage severity as objective marker of pruritus.

Expected output JSON:

{
  "probability_distribution": [
    [
      0.8875784873962402, 0.0006232527666725218, 0.09243898838758469, 0.018707318231463432,
      1.756926576490514e-5, 8.905705908546224e-5, 3.1543219591867455e-8, 0.0005371941952034831,
      7.97201028035488e-6, 2.172054962557013e-7
    ]
  ],
  "continuous_score": 0.2459650261521702
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/excoriation-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Excoriation Intensity Quantification

Configuration File: excoriation_intensity_config.json

Lichenification Intensity Quantification

Purpose: Quantifies skin thickening with accentuated markings indicating chronic disease.

Expected output JSON:

{
  "probability_distribution": [
    [
      0.00338586769066751, 6.331619601951388e-7, 9.158335160464048e-5, 0.05288663133978844,
      0.006485740654170513, 0.00019502690702211112, 0.004077768884599209, 0.9296616911888123,
      0.0031048504170030355, 0.00011023339175153524
    ]
  ],
  "continuous_score": 6.7436911465273965
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/lichenification-classifier/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Lichenification Intensity Quantification

Configuration File: lichenification_intensity_config.json

Wound Characteristic Assessment Models

Model Classification: 🔬 Clinical Model

Purpose: Comprehensive multi-task model providing 26 outputs for wound staging and morphological characterization.

Input Specifications

Parameter	Specification
Type	Single wound image
Format	RGB, uint8 (JPEG, PNG)
Resolution	Variable; resized to model input dimensions
Requirements	Clear view of wound bed and surrounding tissue

Output Specifications

The model produces three categories of outputs:

A. Expected output JSON for Wound Stage (Categorical 0-4)

{
  "predicted_stage": 4.0,
  "probability_distribution": [
    [
      1.714014707943079e-8, 1.1291296431181763e-7, 2.7153060955242836e-7, 0.00023115635849535465,
      0.9997685551643372
    ]
  ]
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/wound-stage-classifier/case-001/image.jpg

B. Expected output JSON for Wound Intensity Score (Regression 0-19)

{
  "score": 10.859453201293945
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/awosi-classifier/case-001/image.jpg

C. Expected output JSON for Binary Wound Characteristics (22 outputs)

Affected_tissues:Bone

{
  "present": false,
  "probability": 1.1421862922134096e-7
}

Affected_tissues:Dermis-epidermis

{
  "present": false,
  "probability": 0.09544984996318817
}

Affected_tissues:Intact

{
  "present": false,
  "probability": 2.890120924803341e-7
}

Affected_tissues:Muscle

{
  "present": false,
  "probability": 1.3682869393960573e-5
}

Affected_tissues:Subcutaneous

{
  "present": false,
  "probability": 0.16069117188453674
}

Biofilm_tissue

{
  "present": true,
  "probability": 0.9993616938591003
}

Borders:Damaged

{
  "present": false,
  "probability": 1.113098824134795e-5
}

Borders:Delimited

{
  "present": true,
  "probability": 0.9999657869338989
}

Borders:Diffused

{
  "present": false,
  "probability": 0.00018668640404939651
}

Borders:Indistinguishable

{
  "present": false,
  "probability": 1.2808607152692275e-6
}

Borders:Thickened

{
  "present": false,
  "probability": 1.1434914313213085e-6
}

Perilesional_erythema

{
  "present": true,
  "probability": 0.8942169547080994
}

Perilesional_maceration

{
  "present": true,
  "probability": 0.5389253497123718
}

Type_exudation:Bloody

{
  "present": false,
  "probability": 2.7391379262553528e-5
}

Type_exudation:Fibrinous

{
  "present": false,
  "probability": 0.03276707977056503
}

Type_exudation:Purulent

{
  "present": true,
  "probability": 0.9628145098686218
}

Type_exudation:Serous

{
  "present": true,
  "probability": 0.9972590208053589
}

Type_tissue_wound_bed:Closed

{
  "present": false,
  "probability": 3.3192190329600635e-8
}

Type_tissue_wound_bed:Epithelial

{
  "present": true,
  "probability": 0.9876804351806641
}

Type_tissue_wound_bed:Granulation

{
  "present": true,
  "probability": 0.920688271522522
}

Type_tissue_wound_bed:Necrotic

{
  "present": false,
  "probability": 8.865220593179401e-7
}

Type_tissue_wound_bed:Slough

{
  "present": true,
  "probability": 0.9999998807907104
}

The original image to generate these outputs can be found in s3://legit-health-plus/test-images/WoundCharacteristicAssessment/1.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Wound Characteristic Assessment

Lesion Quantification Models (Object Detection)

Model Classification: 🔬 Clinical Models

These models detect and count specific lesion types, outputting bounding boxes with confidence scores.

Common Object Detection Specifications

Input Specifications

Parameter	Specification
Type	Single clinical image
Format	RGB, uint8 (JPEG, PNG)
Resolution	Variable; minimum 640×640 recommended

Output Structure

{
  "detections": [
    {
      "class": "lesion_type",
      "confidence": 0.95,
      "box": [120, 85, 180, 145]
    }
  ],
  "counts": {
    "lesion_type": 5
  }
}

Post-processing

Apply Non-Maximum Suppression (NMS)
Filter detections by confidence threshold
Aggregate counts per lesion class

Inflammatory Nodular Lesion Quantification

Purpose: Detects and counts nodules, abscesses, draining tunnels, and non-draining tunnels for IHS4 scoring with oriented bounding boxes.

Detected Classes:

Class	Description	IHS4 Weight
Nodule	Solid inflammatory nodule	×1
Abscess	Pus-filled inflammatory lesion	×2
Non-draining Tunnel	Sinus tract without active drainage	N/A
Draining Tunnel	Sinus tract with active drainage	×4

Output JSON Structure (with oriented bounding boxes in the format xyxyxyxy):

{
  "hs_lesion_detection": {
    "detections": [
      {
        "class": "nodule",
        "confidence": 0.92,
        "box": [
          [185, 348],
          [277, 276],
          [256, 249],
          [164, 321]
        ]
      },
      {
        "class": "abscess",
        "confidence": 0.88,
        "box": [
          [279, 235],
          [291, 165],
          [262, 160],
          [249, 229]
        ]
      },
      {
        "class": "draining_tunnel",
        "confidence": 0.85,
        "box": [
          [255, 266],
          [255, 244],
          [236, 244],
          [236, 266]
        ]
      }
    ],
    "counts": {
      "nodule": 1,
      "abscess": 1,
      "non_draining_tunnel": 0,
      "draining_tunnel": 1
    },
    "ihs4_score": 7
  }
}

IHS4 Calculation: IHS4 = (Nodules × 1) + (Abscesses × 2) + (Draining Tunnels × 4)

Performance: rMAE ≤ 0.45 for each lesion type

IoU threshold for NMS: 0.3

Confidence threshold: 0.3

Configuration File: hs_lesion_detection_config.json

Model's expected results:

s3://legit-health-plus/integration-verification/inflammatory-nodular-lesion-detector/case-001/image.jpg

{
  "hs_lesion_detection": [
    {
      "class": "Draining tunnel",
      "confidence": 0.38,
      "box": [
        [185, 348],
        [277, 276],
        [256, 249],
        [164, 321]
      ]
    }
  ],
  "counts": {
    "Draining tunnel": 1
  },
  "ihs4_score": 4
}

s3://legit-health-plus/integration-verification/inflammatory-nodular-lesion-detector/case-002/image.jpg

{
  "hs_lesion_detection": [
    {
      "class": "Non-draining tunnel",
      "confidence": 0.34,
      "box": [
        [279, 235],
        [291, 165],
        [262, 160],
        [249, 229]
      ]
    }
  ],
  "counts": {
    "Non-draining tunnel": 1
  },
  "ihs4_score": 0
}

s3://legit-health-plus/integration-verification/inflammatory-nodular-lesion-detector/case-003/image.jpg

{
  "hs_lesion_detection": [
    {
      "class": "Nodule",
      "confidence": 0.69,
      "box": [[255, 266], [255, 244], [236, 244], [236, 266]]
    },
    {
      "class": "Nodule",
      "confidence": 0.6,
      "box": [[276, 245], [275, 223], [257, 224], [257, 246]]
    },
    {
      "class": "Nodule",
      "confidence": 0.57,
      "box": [[492, 437], [492, 407], [469, 406], [468, 437]]
    },
    {
      "class": "Nodule",
      "confidence": 0.55,
      "box": [[341, 119], [341, 102], [320, 101], [320, 118]]
    },
    {
      "class": "Nodule",
      "confidence": 0.44,
      "box": [[146, 228], [147, 199], [118, 198], [117, 227]]
    },
    {
      "class": "Nodule",
      "confidence": 0.43,
      "box": [[313, 188], [313, 167], [293, 167], [293, 188]]
    },
    {
      "class": "Nodule",
      "confidence": 0.35,
      "box": [[430, 389], [430, 368], [413, 368], [412, 389]]
    },
    {
      "class": "Nodule",
      "confidence": 0.33,
      "box": [[368, 321], [368, 287], [338, 287], [338, 321]]
    },
    {
      "class": "Nodule",
      "confidence": 0.32,
      "box": [[554, 407], [553, 378], [528, 378], [528, 408]]
    },
    {
      "class": "Nodule",
      "confidence": 0.31,
      "box": [[497, 390], [498, 371], [480, 370], [479, 389]]
    }
  ],
  "counts": {
    "Nodule": 10
  },
  "ihs4_score": 10
}

s3://legit-health-plus/integration-verification/inflammatory-nodular-lesion-detector/case-004/image.jpg

{
  "hs_lesion_detection": [],
  "counts": {},
  "ihs4_score": 0
}

Acneiform Lesion Type Quantification

Purpose: Multi-class detection of acne lesion types for comprehensive severity assessment.

Detected Classes:

Class	Description
Papule	Small solid elevated lesion (<5mm)
Pustule	Pus-filled elevated lesion
Cyst	Deep, pus-filled lesion
Comedone	Open (blackhead) or closed (whitehead)
Nodule	Large solid inflammatory lesion (≥5mm)
Scab	Crusted healing lesion
Spot	Post-inflammatory macule

Output JSON Structure:

{
  "acneiform_lesion_detection": {
    "detections": [
      { "class": "papule", "confidence": 0.89, "box": [50, 60, 70, 80] },
      { "class": "pustule", "confidence": 0.92, "box": [100, 120, 125, 145] },
      { "class": "comedone", "confidence": 0.78, "box": [180, 90, 195, 105] }
    ],
    "counts": {
      "papule": 1,
      "pustule": 1,
      "comedone": 1
    }
  }
}

Performance: rMAE ≤ Expert Inter-observer Variability for each lesion type

IoU threshold for NMS: 0.3

Confidence threshold: 0.15

Configuration File: acneiform_lesion_detection_config.json

Model's expected results:

s3://legit-health-plus/integration-verification/acneiform-detector/case-001/image.jpg

{
  "acneiform_lesion_detection": [
    {
      "class": "Pustule",
      "confidence": 0.43,
      "box": [509, 366, 549, 409]
    },
    {
      "class": "Spot",
      "confidence": 0.42,
      "box": [561, 547, 619, 602]
    },
    {
      "class": "Papule",
      "confidence": 0.38,
      "box": [681, 339, 745, 404]
    },
    {
      "class": "Papule",
      "confidence": 0.35,
      "box": [758, 472, 802, 517]
    },
    {
      "class": "Papule",
      "confidence": 0.35,
      "box": [793, 447, 826, 483]
    },
    {
      "class": "Papule",
      "confidence": 0.34,
      "box": [677, 112, 717, 153]
    },
    {
      "class": "Spot",
      "confidence": 0.32,
      "box": [762, 253, 809, 297]
    },
    {
      "class": "Papule",
      "confidence": 0.31,
      "box": [677, 441, 722, 486]
    },
    {
      "class": "Papule",
      "confidence": 0.3,
      "box": [562, 90, 607, 135]
    },
    {
      "class": "Papule",
      "confidence": 0.29,
      "box": [351, 558, 393, 599]
    },
    {
      "class": "Papule",
      "confidence": 0.28,
      "box": [686, 169, 722, 205]
    },
    {
      "class": "Spot",
      "confidence": 0.25,
      "box": [562, 424, 602, 467]
    },
    {
      "class": "Papule",
      "confidence": 0.22,
      "box": [1129, 342, 1165, 378]
    },
    {
      "class": "Papule",
      "confidence": 0.22,
      "box": [664, 217, 701, 253]
    },
    {
      "class": "Papule",
      "confidence": 0.22,
      "box": [359, 463, 401, 512]
    },
    {
      "class": "Spot",
      "confidence": 0.18,
      "box": [747, 552, 781, 585]
    },
    {
      "class": "Spot",
      "confidence": 0.17,
      "box": [514, 508, 574, 560]
    },
    {
      "class": "Spot",
      "confidence": 0.17,
      "box": [724, 462, 762, 499]
    }
  ],
  "counts": {
    "Pustule": 1,
    "Spot": 6,
    "Papule": 11
  }
}

s3://legit-health-plus/integration-verification/acneiform-detector/case-002/image.jpg

{
  "acneiform_lesion_detection": [
    {
      "class": "Nodule or cyst",
      "confidence": 0.49,
      "box": [250, 160, 284, 192]
    },
    {
      "class": "Pustule",
      "confidence": 0.38,
      "box": [374, 15, 400, 41]
    },
    {
      "class": "Nodule or cyst",
      "confidence": 0.29,
      "box": [59, 148, 118, 220]
    },
    {
      "class": "Scab",
      "confidence": 0.26,
      "box": [399, 177, 426, 207]
    },
    {
      "class": "Nodule or cyst",
      "confidence": 0.21,
      "box": [226, 194, 258, 221]
    }
  ],
  "counts": {
    "Nodule or cyst": 3,
    "Pustule": 1,
    "Scab": 1
  }
}

s3://legit-health-plus/integration-verification/acneiform-detector/case-003/image.jpg

{
  "acneiform_lesion_detection": [
    {
      "class": "Pustule",
      "confidence": 0.51,
      "box": [738, 1167, 935, 1353]
    },
    {
      "class": "Pustule",
      "confidence": 0.49,
      "box": [323, 1289, 406, 1372]
    },
    {
      "class": "Pustule",
      "confidence": 0.35,
      "box": [462, 1321, 493, 1352]
    },
    {
      "class": "Pustule",
      "confidence": 0.28,
      "box": [1270, 714, 1336, 781]
    },
    {
      "class": "Pustule",
      "confidence": 0.28,
      "box": [1231, 1129, 1512, 1360]
    },
    {
      "class": "Papule",
      "confidence": 0.24,
      "box": [451, 1236, 504, 1293]
    },
    {
      "class": "Pustule",
      "confidence": 0.19,
      "box": [1334, 432, 1400, 500]
    },
    {
      "class": "Papule",
      "confidence": 0.17,
      "box": [126, 1156, 177, 1214]
    },
    {
      "class": "Scab",
      "confidence": 0.16,
      "box": [1338, 1202, 1434, 1289]
    },
    {
      "class": "Comedo",
      "confidence": 0.15,
      "box": [940, 529, 973, 562]
    }
  ],
  "counts": {
    "Pustule": 6,
    "Papule": 2,
    "Scab": 1,
    "Comedo": 1
  }
}

Hive Lesion Quantification

Purpose: Detects and counts urticarial wheals for UAS7 scoring.

Output JSON Structure:

{
  "hive_detection": {
    "detections": [
      { "class": "hive", "confidence": 0.85, "box": [80, 100, 150, 170] },
      { "class": "hive", "confidence": 0.78, "box": [200, 150, 280, 230] }
    ],
    "count": 25,
    "uas7_wheal_component": 2
  }
}

UAS7 Wheal Scoring: 0 = none, 1 = <20 wheals, 2 = 20-50 wheals, 3 = >50 wheals

Performance: mAP@50 ≥ 0.56; rMAE ≤ Expert Inter-observer Variability

IoU threshold for NMS: 0.3

Confidence threshold: 0.2

Configuration File: hive_detection_config.json

Model's expected results:

s3://legit-health-plus/test-images/HiveLesionQuantification/1.jpg

{
  "hive_detection": [
    {
      "class": "Hive",
      "confidence": 0.48,
      "box": [81, 381, 129, 438]
    },
    {
      "class": "Hive",
      "confidence": 0.31,
      "box": [571, 161, 657, 231]
    },
    {
      "class": "Hive",
      "confidence": 0.3,
      "box": [1167, 334, 1203, 377]
    },
    {
      "class": "Hive",
      "confidence": 0.3,
      "box": [1084, 358, 1166, 435]
    },
    {
      "class": "Hive",
      "confidence": 0.23,
      "box": [276, 434, 406, 559]
    },
    {
      "class": "Hive",
      "confidence": 0.21,
      "box": [1070, 429, 1151, 507]
    }
  ],
  "counts": {
    "Hive": 6
  },
  "uas7_wheal_component": 1
}

s3://legit-health-plus/test-images/HiveLesionQuantification/2.jpg

{
  "hive_detection": [
    { "class": "Hive", "confidence": 0.79, "box": [232, 718, 291, 772] },
    { "class": "Hive", "confidence": 0.7, "box": [119, 391, 172, 447] },
    { "class": "Hive", "confidence": 0.69, "box": [190, 738, 227, 771] },
    { "class": "Hive", "confidence": 0.65, "box": [316, 724, 351, 755] },
    { "class": "Hive", "confidence": 0.64, "box": [327, 590, 364, 622] },
    { "class": "Hive", "confidence": 0.63, "box": [72, 264, 114, 309] },
    { "class": "Hive", "confidence": 0.62, "box": [250, 160, 288, 196] },
    { "class": "Hive", "confidence": 0.61, "box": [173, 592, 205, 621] },
    { "class": "Hive", "confidence": 0.58, "box": [62, 681, 115, 752] },
    { "class": "Hive", "confidence": 0.58, "box": [293, 134, 347, 179] },
    { "class": "Hive", "confidence": 0.58, "box": [312, 407, 343, 439] },
    { "class": "Hive", "confidence": 0.58, "box": [435, 708, 469, 740] },
    { "class": "Hive", "confidence": 0.57, "box": [51, 167, 96, 221] },
    { "class": "Hive", "confidence": 0.57, "box": [256, 489, 290, 523] },
    { "class": "Hive", "confidence": 0.56, "box": [225, 569, 284, 618] },
    { "class": "Hive", "confidence": 0.54, "box": [390, 249, 439, 299] },
    { "class": "Hive", "confidence": 0.53, "box": [132, 278, 229, 380] },
    { "class": "Hive", "confidence": 0.52, "box": [96, 548, 144, 602] },
    { "class": "Hive", "confidence": 0.5, "box": [120, 195, 217, 270] },
    { "class": "Hive", "confidence": 0.5, "box": [151, 776, 184, 799] },
    { "class": "Hive", "confidence": 0.48, "box": [262, 695, 291, 721] },
    { "class": "Hive", "confidence": 0.48, "box": [187, 539, 217, 569] },
    { "class": "Hive", "confidence": 0.46, "box": [359, 424, 441, 489] },
    { "class": "Hive", "confidence": 0.46, "box": [135, 672, 171, 706] },
    { "class": "Hive", "confidence": 0.45, "box": [205, 604, 233, 631] },
    { "class": "Hive", "confidence": 0.45, "box": [118, 609, 155, 649] },
    { "class": "Hive", "confidence": 0.41, "box": [324, 243, 347, 268] },
    { "class": "Hive", "confidence": 0.38, "box": [246, 421, 278, 452] },
    { "class": "Hive", "confidence": 0.29, "box": [177, 371, 208, 402] },
    { "class": "Hive", "confidence": 0.29, "box": [225, 280, 247, 302] },
    { "class": "Hive", "confidence": 0.28, "box": [225, 187, 257, 220] },
    { "class": "Hive", "confidence": 0.26, "box": [311, 759, 339, 784] },
    { "class": "Hive", "confidence": 0.26, "box": [340, 754, 369, 781] },
    { "class": "Hive", "confidence": 0.25, "box": [287, 372, 319, 401] },
    { "class": "Hive", "confidence": 0.25, "box": [88, 488, 120, 518] },
    { "class": "Hive", "confidence": 0.24, "box": [371, 226, 399, 252] },
    { "class": "Hive", "confidence": 0.23, "box": [450, 244, 472, 264] },
    { "class": "Hive", "confidence": 0.22, "box": [148, 155, 187, 194] },
    { "class": "Hive", "confidence": 0.22, "box": [271, 627, 299, 661] },
    { "class": "Hive", "confidence": 0.21, "box": [229, 360, 254, 389] },
    { "class": "Hive", "confidence": 0.2, "box": [405, 726, 430, 752] },
    { "class": "Hive", "confidence": 0.2, "box": [258, 532, 284, 559] }
  ],
  "counts": {
    "Hive": 42
  },
  "uas7_wheal_component": 2
}

s3://legit-health-plus/test-images/HiveLesionQuantification/3.jpg

{
  "hive_detection": [
    { "class": "Hive", "confidence": 0.73, "box": [1469, 1759, 1762, 2070] },
    { "class": "Hive", "confidence": 0.72, "box": [1087, 815, 1860, 1726] },
    { "class": "Hive", "confidence": 0.7, "box": [307, 1290, 792, 1729] },
    { "class": "Hive", "confidence": 0.65, "box": [225, 361, 581, 714] },
    { "class": "Hive", "confidence": 0.64, "box": [2292, 1523, 2457, 1731] },
    { "class": "Hive", "confidence": 0.62, "box": [160, 1012, 431, 1309] },
    { "class": "Hive", "confidence": 0.61, "box": [1575, 639, 2029, 1168] },
    { "class": "Hive", "confidence": 0.6, "box": [413, 2077, 745, 2443] },
    { "class": "Hive", "confidence": 0.58, "box": [567, 141, 1104, 612] },
    { "class": "Hive", "confidence": 0.57, "box": [180, 644, 453, 959] },
    { "class": "Hive", "confidence": 0.54, "box": [1093, 2234, 1306, 2442] },
    { "class": "Hive", "confidence": 0.54, "box": [1750, 1078, 2451, 1796] },
    { "class": "Hive", "confidence": 0.53, "box": [1972, 1703, 2376, 2143] },
    { "class": "Hive", "confidence": 0.51, "box": [408, 1405, 1507, 2349] },
    { "class": "Hive", "confidence": 0.51, "box": [2309, 0, 2557, 158] },
    { "class": "Hive", "confidence": 0.5, "box": [2117, 654, 2472, 1017] },
    { "class": "Hive", "confidence": 0.49, "box": [2301, 1595, 3223, 2437] },
    { "class": "Hive", "confidence": 0.48, "box": [0, 768, 128, 1038] },
    { "class": "Hive", "confidence": 0.45, "box": [1525, 0, 1877, 264] },
    { "class": "Hive", "confidence": 0.44, "box": [1065, 1482, 1421, 1860] },
    { "class": "Hive", "confidence": 0.43, "box": [1312, 0, 1581, 196] },
    { "class": "Hive", "confidence": 0.43, "box": [2927, 0, 3258, 223] },
    { "class": "Hive", "confidence": 0.43, "box": [2855, 182, 3198, 532] },
    { "class": "Hive", "confidence": 0.4, "box": [2303, 72, 2595, 377] },
    { "class": "Hive", "confidence": 0.4, "box": [1995, 672, 2140, 837] },
    { "class": "Hive", "confidence": 0.36, "box": [804, 1378, 1195, 1772] },
    { "class": "Hive", "confidence": 0.36, "box": [1780, 73, 2435, 634] },
    { "class": "Hive", "confidence": 0.36, "box": [1275, 80, 1953, 532] },
    { "class": "Hive", "confidence": 0.35, "box": [784, 520, 1370, 1033] },
    { "class": "Hive", "confidence": 0.35, "box": [507, 827, 953, 1343] },
    { "class": "Hive", "confidence": 0.33, "box": [887, 132, 1162, 409] },
    { "class": "Hive", "confidence": 0.33, "box": [0, 1298, 145, 1665] },
    { "class": "Hive", "confidence": 0.32, "box": [6, 1045, 215, 1263] },
    { "class": "Hive", "confidence": 0.32, "box": [1445, 405, 1924, 749] },
    { "class": "Hive", "confidence": 0.32, "box": [1673, 1808, 2000, 2097] },
    { "class": "Hive", "confidence": 0.31, "box": [1049, 38, 1311, 294] },
    { "class": "Hive", "confidence": 0.3, "box": [1777, 2053, 2518, 2447] },
    { "class": "Hive", "confidence": 0.29, "box": [1167, 1976, 1514, 2394] },
    { "class": "Hive", "confidence": 0.29, "box": [2620, 1189, 3022, 1605] },
    { "class": "Hive", "confidence": 0.29, "box": [1732, 0, 2126, 152] },
    { "class": "Hive", "confidence": 0.29, "box": [2561, 4, 2915, 389] },
    { "class": "Hive", "confidence": 0.26, "box": [2165, 1094, 2488, 1417] },
    { "class": "Hive", "confidence": 0.26, "box": [2380, 836, 3167, 1759] },
    { "class": "Hive", "confidence": 0.26, "box": [103, 44, 556, 291] },
    { "class": "Hive", "confidence": 0.25, "box": [2386, 848, 2828, 1319] },
    { "class": "Hive", "confidence": 0.24, "box": [730, 0, 1028, 151] },
    { "class": "Hive", "confidence": 0.24, "box": [2210, 376, 2707, 927] },
    { "class": "Hive", "confidence": 0.22, "box": [2456, 311, 3257, 1233] },
    { "class": "Hive", "confidence": 0.2, "box": [2041, 439, 2339, 751] }
  ],
  "count": 49,
  "uas7_wheal_component": 2
}

s3://legit-health-plus/test-images/HiveLesionQuantification/4.jpg

{
  "hs_lesion_detection": [
    { "class": "Hive", "confidence": 0.77, "box": [894, 98, 948, 145] },
    { "class": "Hive", "confidence": 0.71, "box": [713, 271, 770, 325] },
    { "class": "Hive", "confidence": 0.69, "box": [602, 278, 645, 324] },
    { "class": "Hive", "confidence": 0.67, "box": [795, 172, 849, 235] },
    { "class": "Hive", "confidence": 0.65, "box": [795, 126, 842, 178] },
    { "class": "Hive", "confidence": 0.65, "box": [399, 328, 446, 377] },
    { "class": "Hive", "confidence": 0.64, "box": [664, 223, 716, 273] },
    { "class": "Hive", "confidence": 0.64, "box": [147, 609, 199, 650] },
    { "class": "Hive", "confidence": 0.63, "box": [749, 75, 812, 135] },
    { "class": "Hive", "confidence": 0.63, "box": [667, 157, 722, 219] },
    { "class": "Hive", "confidence": 0.61, "box": [912, 179, 961, 227] },
    { "class": "Hive", "confidence": 0.6, "box": [713, 135, 766, 197] },
    { "class": "Hive", "confidence": 0.6, "box": [821, 72, 875, 118] },
    { "class": "Hive", "confidence": 0.59, "box": [371, 209, 461, 292] },
    { "class": "Hive", "confidence": 0.58, "box": [816, 320, 856, 357] },
    { "class": "Hive", "confidence": 0.58, "box": [749, 219, 798, 265] },
    { "class": "Hive", "confidence": 0.58, "box": [621, 224, 660, 267] },
    { "class": "Hive", "confidence": 0.57, "box": [326, 227, 405, 316] },
    { "class": "Hive", "confidence": 0.55, "box": [340, 309, 388, 351] },
    { "class": "Hive", "confidence": 0.55, "box": [349, 352, 399, 396] },
    { "class": "Hive", "confidence": 0.54, "box": [602, 148, 639, 193] },
    { "class": "Hive", "confidence": 0.54, "box": [705, 197, 750, 244] },
    { "class": "Hive", "confidence": 0.54, "box": [822, 272, 870, 323] },
    { "class": "Hive", "confidence": 0.53, "box": [850, 114, 891, 150] },
    { "class": "Hive", "confidence": 0.5, "box": [628, 310, 670, 354] },
    { "class": "Hive", "confidence": 0.5, "box": [764, 42, 814, 82] },
    { "class": "Hive", "confidence": 0.49, "box": [834, 219, 865, 250] },
    { "class": "Hive", "confidence": 0.48, "box": [670, 96, 708, 128] },
    { "class": "Hive", "confidence": 0.48, "box": [783, 268, 820, 306] },
    { "class": "Hive", "confidence": 0.47, "box": [763, 145, 793, 174] },
    { "class": "Hive", "confidence": 0.47, "box": [215, 206, 257, 244] },
    { "class": "Hive", "confidence": 0.47, "box": [890, 149, 933, 188] },
    { "class": "Hive", "confidence": 0.46, "box": [856, 248, 894, 280] },
    { "class": "Hive", "confidence": 0.46, "box": [867, 205, 906, 240] },
    { "class": "Hive", "confidence": 0.45, "box": [1056, 258, 1107, 304] },
    { "class": "Hive", "confidence": 0.45, "box": [680, 57, 721, 95] },
    { "class": "Hive", "confidence": 0.44, "box": [385, 289, 426, 328] },
    { "class": "Hive", "confidence": 0.44, "box": [1011, 190, 1057, 237] },
    { "class": "Hive", "confidence": 0.42, "box": [889, 58, 937, 98] },
    { "class": "Hive", "confidence": 0.41, "box": [370, 638, 409, 671] },
    { "class": "Hive", "confidence": 0.41, "box": [814, 243, 853, 281] },
    { "class": "Hive", "confidence": 0.4, "box": [683, 324, 718, 359] },
    { "class": "Hive", "confidence": 0.4, "box": [844, 156, 881, 197] },
    { "class": "Hive", "confidence": 0.4, "box": [634, 150, 674, 201] },
    { "class": "Hive", "confidence": 0.4, "box": [1048, 215, 1096, 259] },
    { "class": "Hive", "confidence": 0.39, "box": [633, 94, 672, 129] },
    { "class": "Hive", "confidence": 0.38, "box": [692, 118, 732, 156] },
    { "class": "Hive", "confidence": 0.36, "box": [931, 489, 971, 524] },
    { "class": "Hive", "confidence": 0.35, "box": [171, 555, 202, 586] },
    { "class": "Hive", "confidence": 0.35, "box": [446, 216, 488, 256] },
    { "class": "Hive", "confidence": 0.34, "box": [713, 74, 748, 107] },
    { "class": "Hive", "confidence": 0.34, "box": [186, 187, 223, 222] },
    { "class": "Hive", "confidence": 0.34, "box": [752, 187, 784, 218] },
    { "class": "Hive", "confidence": 0.34, "box": [597, 375, 623, 413] },
    { "class": "Hive", "confidence": 0.34, "box": [921, 426, 964, 461] },
    { "class": "Hive", "confidence": 0.33, "box": [768, 299, 807, 334] },
    { "class": "Hive", "confidence": 0.33, "box": [397, 481, 435, 519] },
    { "class": "Hive", "confidence": 0.32, "box": [906, 386, 954, 428] },
    { "class": "Hive", "confidence": 0.31, "box": [474, 250, 504, 281] },
    { "class": "Hive", "confidence": 0.31, "box": [507, 243, 542, 283] },
    { "class": "Hive", "confidence": 0.28, "box": [674, 278, 711, 313] },
    { "class": "Hive", "confidence": 0.28, "box": [651, 339, 681, 370] },
    { "class": "Hive", "confidence": 0.28, "box": [374, 399, 414, 440] },
    { "class": "Hive", "confidence": 0.28, "box": [902, 239, 962, 300] },
    { "class": "Hive", "confidence": 0.26, "box": [802, 404, 837, 439] },
    { "class": "Hive", "confidence": 0.26, "box": [844, 348, 882, 384] },
    { "class": "Hive", "confidence": 0.26, "box": [192, 143, 226, 178] },
    { "class": "Hive", "confidence": 0.25, "box": [486, 177, 531, 224] },
    { "class": "Hive", "confidence": 0.25, "box": [412, 421, 442, 449] },
    { "class": "Hive", "confidence": 0.25, "box": [149, 583, 183, 612] },
    { "class": "Hive", "confidence": 0.24, "box": [1020, 247, 1059, 291] },
    { "class": "Hive", "confidence": 0.23, "box": [823, 499, 866, 540] },
    { "class": "Hive", "confidence": 0.23, "box": [1146, 334, 1199, 395] },
    { "class": "Hive", "confidence": 0.22, "box": [883, 353, 923, 393] },
    { "class": "Hive", "confidence": 0.22, "box": [389, 452, 427, 489] },
    { "class": "Hive", "confidence": 0.21, "box": [352, 205, 390, 236] },
    { "class": "Hive", "confidence": 0.21, "box": [900, 640, 958, 671] },
    { "class": "Hive", "confidence": 0.21, "box": [102, 563, 166, 616] }
  ],
  "count": 78,
  "uas7_wheal_component": 3
}

s3://legit-health-plus/test-images/HiveLesionQuantification/5.jpg

{
  "hs_lesion_detection": [
    { "class": "Hive", "confidence": 0.83, "box": [228, 418, 265, 451] },
    { "class": "Hive", "confidence": 0.83, "box": [262, 434, 305, 472] },
    { "class": "Hive", "confidence": 0.83, "box": [230, 284, 275, 325] },
    { "class": "Hive", "confidence": 0.83, "box": [188, 457, 234, 494] },
    { "class": "Hive", "confidence": 0.83, "box": [258, 383, 307, 427] },
    { "class": "Hive", "confidence": 0.82, "box": [266, 307, 312, 348] },
    { "class": "Hive", "confidence": 0.82, "box": [249, 109, 315, 167] },
    { "class": "Hive", "confidence": 0.81, "box": [151, 412, 191, 445] },
    { "class": "Hive", "confidence": 0.81, "box": [236, 471, 278, 499] },
    { "class": "Hive", "confidence": 0.81, "box": [165, 138, 222, 201] },
    { "class": "Hive", "confidence": 0.8, "box": [292, 396, 346, 445] },
    { "class": "Hive", "confidence": 0.8, "box": [310, 180, 349, 214] },
    { "class": "Hive", "confidence": 0.8, "box": [105, 160, 152, 201] },
    { "class": "Hive", "confidence": 0.8, "box": [168, 85, 219, 132] },
    { "class": "Hive", "confidence": 0.78, "box": [146, 326, 184, 372] },
    { "class": "Hive", "confidence": 0.78, "box": [379, 165, 425, 215] },
    { "class": "Hive", "confidence": 0.78, "box": [216, 168, 258, 208] },
    { "class": "Hive", "confidence": 0.78, "box": [310, 487, 347, 514] },
    { "class": "Hive", "confidence": 0.77, "box": [191, 373, 230, 406] },
    { "class": "Hive", "confidence": 0.77, "box": [403, 120, 442, 159] },
    { "class": "Hive", "confidence": 0.76, "box": [104, 25, 144, 70] },
    { "class": "Hive", "confidence": 0.75, "box": [127, 298, 161, 329] },
    { "class": "Hive", "confidence": 0.75, "box": [132, 430, 165, 461] },
    { "class": "Hive", "confidence": 0.75, "box": [168, 379, 202, 410] },
    { "class": "Hive", "confidence": 0.73, "box": [113, 338, 145, 370] },
    { "class": "Hive", "confidence": 0.72, "box": [449, 81, 479, 128] },
    { "class": "Hive", "confidence": 0.72, "box": [53, 188, 85, 222] },
    { "class": "Hive", "confidence": 0.71, "box": [253, 177, 302, 226] },
    { "class": "Hive", "confidence": 0.71, "box": [170, 501, 241, 545] },
    { "class": "Hive", "confidence": 0.7, "box": [283, 38, 387, 165] },
    { "class": "Hive", "confidence": 0.7, "box": [48, 160, 78, 195] },
    { "class": "Hive", "confidence": 0.69, "box": [240, 10, 300, 122] },
    { "class": "Hive", "confidence": 0.69, "box": [45, 214, 130, 305] },
    { "class": "Hive", "confidence": 0.67, "box": [374, 428, 396, 454] },
    { "class": "Hive", "confidence": 0.67, "box": [330, 224, 360, 254] },
    { "class": "Hive", "confidence": 0.67, "box": [368, 216, 409, 263] },
    { "class": "Hive", "confidence": 0.67, "box": [37, 87, 61, 118] },
    { "class": "Hive", "confidence": 0.66, "box": [292, 524, 342, 546] },
    { "class": "Hive", "confidence": 0.66, "box": [339, 495, 381, 527] },
    { "class": "Hive", "confidence": 0.65, "box": [237, 512, 264, 533] },
    { "class": "Hive", "confidence": 0.64, "box": [324, 405, 379, 456] },
    { "class": "Hive", "confidence": 0.62, "box": [86, 76, 124, 119] },
    { "class": "Hive", "confidence": 0.62, "box": [358, 273, 408, 313] },
    { "class": "Hive", "confidence": 0.62, "box": [320, 0, 413, 43] },
    { "class": "Hive", "confidence": 0.61, "box": [416, 316, 442, 350] },
    { "class": "Hive", "confidence": 0.58, "box": [236, 343, 282, 384] },
    { "class": "Hive", "confidence": 0.58, "box": [323, 461, 349, 483] },
    { "class": "Hive", "confidence": 0.57, "box": [103, 381, 151, 427] },
    { "class": "Hive", "confidence": 0.56, "box": [384, 7, 451, 77] },
    { "class": "Hive", "confidence": 0.49, "box": [355, 356, 391, 391] },
    { "class": "Hive", "confidence": 0.49, "box": [385, 401, 406, 425] },
    { "class": "Hive", "confidence": 0.48, "box": [244, 446, 263, 462] },
    { "class": "Hive", "confidence": 0.47, "box": [48, 117, 69, 140] },
    { "class": "Hive", "confidence": 0.44, "box": [59, 8, 101, 68] },
    { "class": "Hive", "confidence": 0.41, "box": [215, 341, 251, 374] },
    { "class": "Hive", "confidence": 0.31, "box": [150, 370, 168, 387] },
    { "class": "Hive", "confidence": 0.3, "box": [101, 286, 129, 314] },
    { "class": "Hive", "confidence": 0.25, "box": [371, 368, 400, 402] },
    { "class": "Hive", "confidence": 0.25, "box": [406, 395, 426, 430] }
  ],
  "count": 59,
  "uas7_wheal_component": 3
}

s3://legit-health-plus/test-images/HiveLesionQuantification/6.jpg

{
  "hs_lesion_detection": [
    {
      "class": "Hive",
      "confidence": 0.76,
      "box": [14, 67, 237, 389]
    },
    {
      "class": "Hive",
      "confidence": 0.41,
      "box": [162, 160, 396, 390]
    },
    {
      "class": "Hive",
      "confidence": 0.35,
      "box": [251, 64, 380, 181]
    },
    {
      "class": "Hive",
      "confidence": 0.34,
      "box": [382, 209, 537, 421]
    },
    {
      "class": "Hive",
      "confidence": 0.23,
      "box": [479, 237, 595, 435]
    }
  ],
  "count": 5,
  "uas7_wheal_component": 1
}

Surface Area Quantification Models (Segmentation)

Model Classification: 🔬 Clinical Models

These models perform pixel-wise segmentation to quantify affected surface areas.

Common Segmentation Specifications

Input Specifications

Parameter	Specification
Type	Single clinical image
Format	RGB, uint8 (JPEG, PNG)
Resolution	Variable; original resolution preserved for output

Output Structure

{
  "segmentation": {
    "mask": "base64_encoded_png",
    "probability_map": "base64_encoded_png",
    "percentage_affected": 15.7,
    "pixel_count": 45230,
    "total_roi_pixels": 288000
  }
}

Post-processing

Apply sigmoid activation to obtain probability map
Threshold at 0.5 to generate binary mask
Calculate percentage: (affected_pixels / total_roi_pixels) × 100

Wound Surface Quantification

Purpose: 7-class segmentation for comprehensive wound tissue analysis.

Segmentation Classes:

Class	Description
Wound Bed	Total wound area
Granulation	Healthy healing tissue
Biofilm/Slough	Devitalized tissue
Necrosis	Non-viable tissue
Maceration	Periwound moisture damage
Orthopedic Material	Exposed hardware
Bone/Cartilage/Tendon	Exposed deep structures

Expected output JSON:

Wound Bed

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 3.072578463203463
}

Biofilm/Slough

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 13.472620646158854
}

Granulation

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 10.111006879299563
}

Necrosis

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 4.453454000647878
}

Maceration

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 4.4342041015625
}

Orthopedic Material

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 0.801688345091123
}

Bone/Cartilage/Tendon

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 1.1587196607037877
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/wound-***/case-xyz/image.jpg

Configuration File: wound_surface_segmentation_config.json

Erythema Surface Quantification

Purpose: Binary segmentation for perilesional and wound bed erythema extent.

Expected output JSON:

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 17.151666666666667
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/erythema-segmenter/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Erythema Surface Quantification

Configuration File: erythema_surface_config.json

Hair Loss Surface Quantification

Purpose: 3-class segmentation (Hair, No Hair, Non-Scalp) for alopecia assessment and SALT scoring.

Expected output JSON Structure:

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": [
    "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
    "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
    "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ..."
  ],
  "percentages": [70.08219203995158, 25.789289497578693, 4.128518462469734],
  "alopecia_percentages": 13.799535273382546
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/hair-loss-segmenter/case-001/image.jpg

Calculation: Hair Loss % = (No Hair pixels) / (Hair + No Hair pixels) × 100

Performance: Specified in R-TF-028-005 AI Development Report: Hair Loss Surface Quantification

Configuration File: hair_loss_surface_config.json

Nail Lesion Surface Quantification

Purpose: Multi-class segmentation (Healthy Nail, Lesion, Background) for the assessment of nail lesion extent.

Output JSON Structure:

{
  "image": "base64_png",
  "probability": "base64_png",
  "background": "base64_png",
  "healthy nail": "base64_png",
  "nail lesion": "base64_png"
}

Performance Summary

Metric	Result	Success Criterion
IoU (overall nail segmentation)	0.8900 (95% CI: [0.8712-0.9061])	≥ 0.80
IoU (nail lesion segmentation)	0.8195 (95% CI: [0.7934-0.8418])	≥ 0.70

Configuration File

Configuration File: nail_lesion_surface_config.json

{
  "architecture": "unet",
  "backbone": "resnet101",
  "classes": ["background", "healthy nail", "nail lesion"],
  "img_size": 480,
  "img_stats": {
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225]
  },
  "threshold": 0.5
}

Expected Results

s3://legit-health-plus/integration-verification/nail-lesion-segmenter/case-001/image.jpg

{
  "image": "b'/9j/4AAQSkZJRgABAQ...",
  "probability": "b'/9j/4AAQSkZJRgABAQAAAQABAAD/...",
  "background": "b'/9j/4AAQSkZJRgABAQAAAQABAAD/...",
  "healthy nail": "b'/9j/4AAQSkZJRgABAQAAAQABAAD/...",
  "nail lesion": "b'/9j/4AAQSkZJRgABAQAAAQABAAD/..."
}

Hypopigmentation/Depigmentation Surface Quantification

Purpose: Binary segmentation for vitiligo and pigmentary loss quantification (VASI scoring).

Expected output JSON:

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 18.5432266198044
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/hypopigmentation-segmenter/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Hypopigmentation/Depigmentation Surface Quantification

Configuration File: hypopigmentation_surface_config.json

Hyperpigmentation Surface Quantification

Purpose: Binary segmentation for melasma and PIH quantification (MASI scoring).

Expected output JSON:

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 34.57489013671875
}

The original image to generate these outputs can be found in s3://legit-health-plus/integration-verification/hyperpigmentation-segmenter/case-001/image.jpg

Performance: Specified in R-TF-028-005 AI Development Report: Hyperpigmentation Surface Quantification

Configuration File: hyperpigmentation_surface_config.json

Pattern Identification Models

Model Classification: 🔬 Clinical Models

These models classify disease patterns and severity stages.

Follicular and Inflammatory Pattern Identification

Purpose: Multi-class classification for HS phenotype identification (Martorell classification).

Input Specifications

Parameter	Specification
Type	Single clinical image of follicular/inflammatory lesions
Format	RGB, uint8 (JPEG, PNG)

Output Specifications

Phenotype Classes:

The model predicts the probability for the following phenotypes

Phenotype	Description
Follicular	Lesions from hair follicles, comedones, sinus tracts
Inflammatory	Sudden-onset abscesses without prominent follicular lesions

Output classes:

The output probabilites of the model are thresholded and aggregated to produce one of the following output classes:

Phenotype	Description
Nothing	Follicular = False, Inflammatory = False
Follicular	Follicular = True, Inflammatory = False
Inflammatory	Follicular = False, Inflammatory = True
Mixed	Follicular = True, Inflammatory = True

Output JSON Structure:

{
  "hs_phenotype": {
    "predicted_phenotype": "Inflammatory",
    "probability_distribution": {
      "follicular": 0.15,
      "inflammatory": 0.72
    }
  }
}

Performance Summary

Metric	Result	Success Criterion
BACC	0.6837 (95% CI: [0.6287-0.7398])	≥ 0.65
Average F1 score	0.6976 (95% CI: [0.6457-0.7526])	≥ 0.65

Configuration File

Configuration File: follicular_inflammatory_pattern_config.json

{
  "architecture": "convnextv2_base.fcmae_ft_in1k",
  "hidden_features": 512,
  "classes": ["Nothing", "Follicular", "Inflammatory", "Mixed"],
  "img_size": 480,
  "img_stats": {
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225]
  },
  "threshold": 0.5
}

Expected Results

s3://legit-health-plus/integration-verification/follicular-inflammatory-pattern-classifier/case-001/image.jpg

{
  "hs_phenotype": {
    "predicted_phenotype": "Mixed",
    "probability": {
      "Follicular": 0.9934564232826233,
      "Inflammatory": 0.9881375432014465
    }
  }
}

Inflammatory Pattern Identification (Hurley Staging)

Purpose: Multi-task classification model that simultaneously outputs Hurley stage classification and inflammatory activity assessment for inflammatory nodular lesions.

Output Specifications

This multi-task model provides two simultaneous classifications:

Hurley Stage Classification: Four-class categorization (Clear, Stage I, Stage II, Stage III)
Inflammatory Activity Classification: Binary categorization (Inflammatory, No Inflammatory)

Hurley Stage Definitions:

Stage	Description
Clear	No visible inflammatory lesions
Stage I	Single or multiple isolated abscesses without sinus tracts or scarring
Stage II	Recurrent abscesses with sinus tract formation and scarring
Stage III	Diffuse or broad involvement with multiple interconnected sinus tracts and abscesses across entire anatomical area

Inflammatory Activity Definitions:

Status	Description
No Inflammatory	Inactive disease with post-inflammatory changes (scars, fibrotic tracts, comedones without erythema, healed lesions)
Inflammatory	Active disease with erythematous nodules, abscesses, draining sinus tracts, acute flares, signs of acute inflammation

Output JSON Structure:

{
  "hurley_staging": "Stage II",
  "hurley_staging_prob": 0.59,
  "inflammatory_activity": "No inflammatory",
  "inflammatory_activity_prob": 0.93
}

Performance:

Hurley Stage: Accuracy ≥ 40%; Mean Absolute Error (MAE) ≤ 1
Inflammatory Activity: Accuracy ≥ 70%; AUC (ROC) ≥ 0.70

Configuration File: inflammatory_nodular_lesion_pattern_config.json

Model's expected results:

s3://legit-health-plus/integration-verification/inflammatory-pattern-identificator/case-001/image.jpg

{
    "hurley_staging": "Stage III",
    "hurley_staging_prob": 0.87,
    "inflammatory_activity": "No inflammatory",
    "inflammatory_activity_prob": 0.95
}

s3://legit-health-plus/integration-verification/inflammatory-pattern-identificator/case-002/image.jpg

{
    "hurley_staging": "Stage III",
    "hurley_staging_prob": 0.62,
    "inflammatory_activity": "No inflammatory",
    "inflammatory_activity_prob": 0.9
}

s3://legit-health-plus/integration-verification/inflammatory-pattern-identificator/case-003/image.jpg

{
    "hurley_staging": "Stage I",
    "hurley_staging_prob": 0.59,
    "inflammatory_activity": "No inflammatory",
    "inflammatory_activity_prob": 0.98
}

s3://legit-health-plus/integration-verification/inflammatory-pattern-identificator/case-004/image.jpg

{
    "hurley_staging": "Stage II",
    "hurley_staging_prob": 0.55,
    "inflammatory_activity": "No inflammatory",
    "inflammatory_activity_prob": 0.6
}

Hair Follicle Quantification

Purpose: Single-class detection of hair follicle for assessment of hair loss.

Output JSON Structure:

{
  "hair_follicle_detection": {
    "detections": [
      { "class": "follicle", "confidence": 0.89, "box": [50, 60, 70, 80] },
      { "class": "follicle", "confidence": 0.92, "box": [100, 120, 125, 145] },
      { "class": "follicle", "confidence": 0.78, "box": [180, 90, 195, 105] }
    ],
    "counts": 1
  }
}

IoU threshold for NMS: 0.3

Confidence threshold: 0.15

Performance Summary

Metric	Result	Success Criterion
mAP@50	0.8162 (95% CI: [0.7503 - 0.8686])	≥ 0.72

Configuration File

Configuration File: hair_follicle_detection_config.json

Expected Results

s3://legit-health-plus/integration-verification/hair-follicle-detector/case-001/image.jpg

{
  "hair_follicle_detection": {
    "detections": [
      {
        "class": "follicle",
        "confidence": 0.506977,
        "box": [507, 371, 603, 480]
      },
      {
        "class": "follicle",
        "confidence": 0.246984,
        "box": [291, 288, 380, 411]
      },
      {
        "class": "follicle",
        "confidence": 0.160094,
        "box": [0, 6, 45, 159]
      }
    ],
    "counts": 3
  }
}

Non-Clinical Models: Detailed Specifications

Domain Validation

Model Classification: 🛠️ Non-Clinical Model

Purpose: Validates image domain (non-dermatology, skin (clinical), skin (dermoscopic)) for proper routing.

Input Specifications

Parameter	Specification
Type	Any image
Format	RGB, uint8 (JPEG, PNG)

Output Specifications

Output JSON Structure:

{
  "domain_validation": {
    "predicted_domain": "skin_clinical",
    "probability_distribution": {
      "non_dermatology": 0.02,
      "clinical": 0.92,
      "dermoscopic": 0.06
    }
  }
}

Performance Summary

Metric	Value	Criterion
Non-dermatology precision	0.9855 (95% CI: [0.9828-0.9882])	≥ 0.95
Non-dermatology recall	0.9978 (95% CI: [0.9967-0.9988])	≥ 0.90
Clinical f1-score	0.9975 (95% CI: [0.9973-0.9978])	≥ 0.90
Dermoscopic f1-score	0.9950 (95% CI: [0.9942-0.9957])	≥ 0.90
Accuracy	0.9965 (95% CI: [0.9961-0.9969])	≥ 95%
Macro avg f1-score	0.9947 (95% CI: [0.9940-0.9953])	≥ 0.90
Weighted avg f1-score	0.9965 (95% CI: [0.9961-0.9969])	≥ 0.90

Configuration File

Configuration File: domain_validation_config.json

{
  "architecture": "efficientnet_b0.ra_in1k",
  "classes": ["non-dermatology", "clinical", "dermoscopic"],
  "img_size": 224,
  "img_stats": {
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225]
  }
}

Expected Results

s3://legit-health-plus/integration-verification/domain-validation/case-001/image.jpg

{
  "domain_validation": {
    "predicted_domain": "non-dermatology",
    "probability_distribution": {
      "non-dermatology": 1.0,
      "clinical": 2.495789094747707e-21,
      "dermoscopic": 1.781282008838602e-19
    }
  }
}

Dermatology Image Quality Assessment (DIQA)

Model Classification: 🛠️ Non-Clinical Model

Purpose: Assesses visual quality for clinical acceptability.

Output Specifications

The model outputs a logit (a scalar) that is rescaled to the [1, 10] range.

Output JSON Structure:

{
  "image_quality": {
    "overall_score": 7.5
  }
}

Performance Summary

Metric	Result	Success Criterion
Pearson correlation	0.8959 (95% CI: [0.8910-0.9002])	`≥ 0.70`
Spearman correlation	0.9030 (95% CI: [0.8982-0.9071])	`≥ 0.70`

Configuration File

Configuration File: diqa_config.json

{
  "architecture": "efficientnet_b5.sw_in12k",
  "embeddings_size": "768",
  "num_ratings": 10,
  "img_size": 416,
  "img_stats": {
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225]
  },
  "score_scaler": { "mean": 5.782282610281528, "std": 1.773072826496509 }
}

Expected Results

s3://legit-health-plus/integration-verification/diqa/case-001/image.jpg

{
  "image_quality": {
    "overall_score": 2.7366
  }
}

Skin Surface Segmentation

Model Classification: 🛠️ Non-Clinical Model

Purpose: Binary segmentation to isolate skin regions from background.

Output Specifications

Output JSON Structure:

{
  "image": "iVBORw0KGgoAAAANSUhEUgAABRAAAALYCAIAAACL4kmEAAEAAE",
  "mask": "iVBORw0KGgoAAAANSUhEUgAABRAAAALYCAAAAAAh64EPAAAKmk",
  "probability_mask": "iVBORw0KGgoAAAANSUhEUgAABRAAAALYCAAAAAAh64EPAAAhxk",
  "skin_percentage": 0.83
}

Performance: Mean IoU ≥ 0.83; F1-score ≥ 0.84

Configuration File: skin_segmentation_config.json

Output JSON Structure:

s3://legit-health-plus/integration-verification/skin-segmenter/case-001/image.jpg

{'image': 'iVBORw0KGgoAAAANSUhEUgAABRAAAALYCAIAAACL4kmEAAEAAE',
 'mask': 'iVBORw0KGgoAAAANSUhEUgAABRAAAALYCAAAAAAh64EPAAAKmk',
 'probability_mask': 'iVBORw0KGgoAAAANSUhEUgAABRAAAALYCAAAAAAh64EPAAAhxk',
 'skin_percentage': 0.83}

Body Surface Segmentation

Model Classification: 🛠️ Non-Clinical Model

Purpose: Segments body surface for BSA calculations in severity scoring (PASI, EASI).

Output Specifications

Output JSON Structure:

{
  "image": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "mask": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "probability_map": "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQ...",
  "percentage": 53.62202037464488
}

Performance: Specified in R-TF-028-005 AI Development Report: Body Surface Segmentation

Head Detection

Purpose: Object detection for automated head localization to support hair loss quantification and image standardization workflows.

Input Specifications

Parameter	Specification
Type	Single clinical image
Format	RGB, uint8 (JPEG, PNG)
Size	Variable; resized to 480×480 pixels for inference

Output Specifications

Model Classification: 🛠️ Non-Clinical Model

Output JSON Structure:

{
  "detections": [
    {
      "box": [x_min, y_min, x_max, y_max],
      "confidence": 0.95
    }
  ],
  "head_count": 1
}

Bounding Box Format: [x_min, y_min, x_max, y_max] in pixel coordinates

Post-processing:

Confidence threshold: 0.25
Non-maximum suppression (NMS) with IoU threshold: 0.7

Performance: mAP@50 ≥ 0.86

Configuration File: head_detection_config.json

Output JSON Structure:

s3://legit-health-plus/integration-verification/head-detector/case-001/image.jpg

{
  "detections": [
    {
      "confidence": 0.91,
      "box": [51, 28, 183, 200]
    },
    {
      "confidence": 0.87,
      "box": [776, 85, 902, 258]
    },
    {
      "confidence": 0.86,
      "box": [495, 129, 611, 288]
    },
    {
      "confidence": 0.78,
      "box": [978, 154, 1112, 337]
    },
    {
      "confidence": 0.71,
      "box": [638, 43, 751, 208]
    },
    {
      "confidence": 0.67,
      "box": [128, 213, 273, 380]
    },
    {
      "confidence": 0.46,
      "box": [273, 176, 395, 345]
    }
  ],
  "count": 7
}

Technical Integration Details

API Orchestration Logic

The backend service orchestrates model calls as follows. The API caller selects which clinical models to execute via the request payload. All requested models execute independently and all outputs are returned together in a single response:

1. RECEIVE image + requested model list via API endpoint

2. NON-CLINICAL MODELS (always executed)
   ├── Domain Validation → Informational classification (clinical / dermoscopic / non-dermatology)
   └── Image Quality Assessment (DIQA) → Informational quality score (0–10)

3. CLINICAL MODELS (selected by API caller, parallel execution)
   ├── Skin Surface Segmentation (as requested)
   ├── ICD Category Distribution
   │   └── Binary Indicator Derivation (deterministic post-processing)
   ├── Visual Sign Intensity Models (as requested)
   ├── Lesion Detection Models (as requested)
   ├── Surface Quantification Models (as requested)
   └── Pattern Identification Models (as requested)

4. AGGREGATE all outputs into structured JSON response

5. RETURN complete response including non-clinical + clinical outputs (API is stateless)

Key architectural principle: Non-clinical models (Domain Validation, DIQA) execute on every request and their outputs are included in the response as informational metadata. They do not gate, block, or condition the execution of clinical models. If the API caller requests clinical analysis, all requested clinical models execute regardless of the Domain Validation classification or DIQA score. The clinical outputs are always returned alongside the non-clinical outputs, allowing the consuming application to decide how to present or act on the combined information.

Model Execution Order

Priority	Model(s)	Dependency
1	Domain Validation	None (always executed)
1	Image Quality Assessment	None (always executed)
2	Skin Surface Segmentation	None (executes if requested by caller)
2	ICD Category Distribution	None (executes if requested by caller)
2	All other clinical models	None (execute in parallel if requested by caller)
3	Binary Indicators	ICD distribution complete (deterministic derivation)

All models at the same priority level execute in parallel. Non-clinical models (Priority 1) always execute. Clinical models (Priority 2) execute based on the API caller's request. Binary Indicators (Priority 3) are the only true dependency in the pipeline — they are derived deterministically from the ICD Category Distribution output via the documented mapping matrix.

Error Handling

Condition	Action
Domain = non_skin	Domain Validation output included in response as informational metadata; clinical models still execute if requested
Quality < 4	DIQA score included in response as informational metadata; clinical models still execute if requested
Skin ROI < 10%	Segmentation output included in response; clinical models still execute if requested
Model inference failure	Return partial results with error flag for the failed model

Note: The error handling strategy reflects the informational nature of non-clinical model outputs. The API always returns all requested outputs. The consuming application is responsible for interpreting the non-clinical outputs (e.g., presenting a retake recommendation to the user when DIQA is low) and deciding how to present the clinical outputs in context.

Multi-Model Consistency

The Legit.Health Plus device integrates 59 AI models that must produce coherent, clinically consistent outputs when executed together on the same input image. This section documents the design principles, architectural safeguards, and verification mechanisms that ensure multi-model consistency across the orchestration pipeline.

Design Principles for Consistency

1. Common Preprocessing and Informational Non-Clinical Outputs:

All models receive inputs that have been processed through a common, deterministic preprocessing pipeline. This ensures that every model operates on a consistent representation of the same input image:

Domain Validation and DIQA execute on every request and produce informational outputs that are returned alongside clinical model outputs. They do not gate, block, or condition the execution of any clinical model.
Skin Surface Segmentation executes when requested by the API caller and provides a Region of Interest (ROI) mask that other clinical models can use for skin localization, ensuring that models analyzing the same image operate on consistent anatomical regions.
Preprocessing operations (resize, normalization, color space conversion) are applied identically per model category, as specified in the Configuration Files section of this document.

2. Deterministic Execution:

All models are offline (static) — they do not learn or adapt during deployment. This guarantees that:

The same input always produces the same output (given identical preprocessing and runtime configuration).
Model behavior is fully reproducible and verifiable across environments.
No feedback loops exist between models that could introduce instability or drift.

3. Independent Execution with Minimal Dependencies:

All clinical models execute independently — there is no inter-model conditioning or gatekeeping. The only true dependency in the pipeline is the deterministic derivation of Binary Indicators from the ICD Category Distribution output:

Relationship	Description
Domain Validation, DIQA	Always execute; outputs are informational metadata included in the response
Clinical models	Execute independently in parallel based on the API caller's request — no inter-model dependencies
ICD Distribution → Binary Indicators	The only true dependency: Binary Indicators are derived deterministically from the ICD probability distribution via a documented mapping matrix

4. Independent Clinical Model Outputs:

Critically, the clinical models that execute in parallel (Priority 6 in the Model Execution Order) are architecturally independent — each model produces its output based solely on the input image and its own learned parameters, without receiving or being influenced by the outputs of other clinical models. This design ensures that:

A prediction error in one model does not propagate to other models.
Each model can be independently validated, retrained, and updated without affecting the behavior of other models.
The aggregate clinical output is a composition of independently verified components.

Cross-Model Coherence Verification

Although clinical models are architecturally independent, the aggregate output must be clinically coherent. The following verification mechanisms ensure cross-model consistency:

1. Input Consistency Verification:

The Integration Verification Package (see below) ensures that all models receive correctly preprocessed inputs in the production environment. Any preprocessing discrepancy that could cause inconsistency between models is detected during integration verification.

2. Output Schema Validation:

The API response aggregation step (Step 4 in the orchestration flow) validates that all model outputs conform to their expected schemas before the response is returned:

Classification outputs must contain valid probability distributions (sum to 1.0 within tolerance).
Regression outputs must fall within their defined clinical ranges (e.g., 0–9 for intensity scores).
Segmentation masks must have valid dimensions matching the input image.
Detection outputs must have valid bounding box coordinates within image boundaries.

3. Clinical Plausibility Checks:

The following post-processing validation rules are applied to detect clinically implausible output combinations:

Check	Rule	Action on Violation
ICD–Binary Indicator Alignment	Binary indicators must be derivable from the ICD distribution via the documented mapping matrix	Flag inconsistency in response metadata
Segmentation–Detection Coherence	For models that assess the same anatomical region, segmentation mask area and detection count should be directionally consistent	Log warning for post-market monitoring

4. End-to-End Integration Testing:

The Integration Verification Package provides reference test images with expected outputs for all 59 models. By processing the same image through the complete orchestration pipeline, the Software Development team verifies that:

All models execute successfully on the same input.
Outputs are consistent with development-validated expected outputs.
The aggregated response conforms to the documented API schema.
No inter-model interference occurs during parallel execution.

Versioning and Consistency Across Updates

When individual models are retrained or updated (per the AI Model Versioning scheme in GP-028):

The complete Integration Verification Package is regenerated for the updated model.
End-to-end integration testing is repeated with the full 59-model pipeline to confirm that the updated model does not introduce inconsistencies with other models.
Cross-model coherence checks are re-executed as part of the R-TF-028-010 AI V&V Checks.

This multi-model consistency framework ensures that the 59 individually validated AI models collectively produce reliable, coherent, and clinically meaningful outputs when orchestrated within the Legit.Health Plus device.

Integration Verification Package

To ensure that AI models produce identical outputs when integrated into the target software environment as they did during development and validation, this release includes an Integration Verification Package for each model, as required by GP-028.

Purpose

The Integration Verification Package serves to:

Verify Integration Correctness: Confirm that models are correctly integrated without alterations to their inference behavior.
Detect Environment Discrepancies: Identify any differences in preprocessing, numerical precision, or runtime configuration that could affect model outputs.
Ensure Reproducibility: Provide objective evidence that the integrated model produces the same outputs as the validated development model.
Support Regulatory Compliance: Demonstrate traceability between development validation and deployed system verification per IEC 62304 and MDR requirements.

Package Contents

For each AI model in this release, the Integration Verification Package includes:

Reference Test Images

A curated subset of images from the model's test set, stored in a secure, version-controlled location:

Model Category	Storage Location
ICD Category Distribution	`s3://legit-health-plus/integration-verification/icd-distribution/v1.1.0/images/`
Binary Indicators	Uses ICD Category Distribution images
Visual Sign Intensity Models	`s3://legit-health-plus/integration-verification/intensity-models/v1.1.0/images/`
Wound Characteristic Assessment	`s3://legit-health-plus/integration-verification/wound-assessment/v1.1.0/images/`
Lesion Quantification Models	`s3://legit-health-plus/integration-verification/lesion-detection/v1.1.0/images/`
Surface Segmentation Models	`s3://legit-health-plus/integration-verification/segmentation/v1.1.0/images/`
Pattern Identification Models	`s3://legit-health-plus/integration-verification/pattern-models/v1.1.0/images/`
Domain Validation	`s3://legit-health-plus/integration-verification/domain-validation/v1.1.0/images/`
DIQA	`s3://legit-health-plus/integration-verification/diqa/v1.1.0/images/`
Skin Surface Segmentation	`s3://legit-health-plus/integration-verification/skin-segmentation/v1.1.0/images/`
Body Surface Segmentation	`s3://legit-health-plus/integration-verification/body-segmentation/v1.1.0/images/`

Images are representative of the model's input domain (diverse conditions, demographics, imaging modalities) and stored in their original format without additional processing.

Expected Outputs File

A CSV file containing the expected model outputs for each reference image:

Storage location: s3://legit-health-plus/integration-verification/{MODEL_NAME}/v1.1.0/expected_outputs.csv
File schema:

Column	Description
`image_id`	Unique identifier matching the image filename
`expected_output`	The model's expected output (format varies by model type)
`output_type`	Classification probability, regression value, segmentation mask hash, etc.
`confidence`	Model confidence score (where applicable)

Example for ICD Category Distribution:

image_id,expected_output,output_type,confidence
IMG_001.jpg,"[0.72, 0.12, 0.08, 0.04, 0.02, ...]",probability_distribution,0.72
IMG_002.jpg,"[0.65, 0.18, 0.09, 0.05, 0.02, ...]",probability_distribution,0.65

Example for Visual Sign Intensity:

image_id,expected_output,output_type,confidence
IMG_001.jpg,4.52,continuous_score,0.87
IMG_002.jpg,3.85,continuous_score,0.91

Example for Segmentation Models:

image_id,expected_output,output_type,mask_hash
IMG_001.jpg,15.7,percentage_affected,sha256:abc123...
IMG_002.jpg,22.3,percentage_affected,sha256:def456...

Verification Manifest

A JSON manifest file containing metadata for the verification package:

Storage location: s3://legit-health-plus/integration-verification/{MODEL_NAME}/v1.1.0/manifest.json

Example manifest:

{
  "model_name": "icd_distribution",
  "model_version": "1.1.0.0",
  "package_version": "1.0.0",
  "created_timestamp": "2025-11-15T10:30:00Z",
  "created_by": "AI Team",
  "test_image_count": 50,
  "acceptance_criteria": {
    "metric": "probability_match",
    "tolerance": 1e-5,
    "pass_rate_required": 1.0
  },
  "preprocessing": {
    "resize": [384, 384],
    "normalization_mean": [0.485, 0.456, 0.406],
    "normalization_std": [0.229, 0.224, 0.225]
  },
  "runtime_config": {
    "pytorch_version": ">=1.12.0",
    "tta_enabled": true,
    "temperature_scaling": 1.05
  },
  "model_weights_hash": "sha256:e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
}

Acceptance Criteria

The following acceptance criteria apply to integration verification:

Model Type	Metric	Acceptance Criterion
Classification (ICD, Domain, Patterns)	Output probability match	Difference ≤ 1e-5 per element
Binary Indicators	Derived indicator match	Difference ≤ 1e-5 per indicator
Intensity Quantification	Continuous score match	Difference ≤ 1e-4
Object Detection	Bounding box IoU	IoU ≥ 0.9999
Object Detection	Detection count	Exact match
Segmentation	Pixel-wise mask comparison	IoU ≥ 0.9999
Segmentation	Percentage affected	Difference ≤ 0.01%
DIQA	Quality score match	Difference ≤ 1e-4

Overall Pass Rate: 100% of test images must meet the acceptance criteria for the integration to be considered successful.

Verification Procedure

The Software Development team shall execute the following procedure after model integration:

Step 1: Environment Preparation

Configure the integration environment with the same dependencies and versions specified in this Release Report
Download the Integration Verification Package from the designated S3 location
Verify package integrity using manifest checksums:
```
sha256sum -c checksums.txt
```

Step 2: Inference Execution

Process all reference test images through the integrated model
Apply the same preprocessing as specified in the manifest
Record outputs in the same format as the expected outputs file
Document the runtime environment configuration

Step 3: Output Comparison

Compare actual outputs against expected outputs using the specified acceptance criteria
Calculate the match rate for each metric type
Identify and document any discrepancies

Example verification script structure:

import pandas as pd
import numpy as np

def verify_integration(expected_csv, actual_csv, tolerance=1e-5):
    expected = pd.read_csv(expected_csv)
    actual = pd.read_csv(actual_csv)

    results = []
    for idx, row in expected.iterrows():
        expected_output = np.array(eval(row['expected_output']))
        actual_output = np.array(eval(actual.loc[idx, 'expected_output']))

        diff = np.max(np.abs(expected_output - actual_output))
        passed = diff <= tolerance

        results.append({
            'image_id': row['image_id'],
            'max_difference': diff,
            'passed': passed
        })

    return pd.DataFrame(results)

Step 4: Results Documentation

Generate an Integration Verification Report documenting:

Test execution date and environment details
Pass/fail status for each test image
Overall pass rate and compliance with acceptance criteria
Any deviations and their root cause analysis

This report becomes part of the software verification evidence per IEC 62304.

Failure Handling

If integration verification fails:

Root Cause Analysis: Investigate the source of discrepancy:
- Preprocessing differences (resize algorithm, interpolation method)
- Numerical precision (float32 vs float64)
- Library version differences (ONNX runtime, numpy)
- Post-processing implementation differences
Corrective Action: Implement necessary corrections to achieve output equivalence
Re-verification: Repeat the verification procedure after corrections
Escalation: If equivalence cannot be achieved, escalate to the AI Team for joint investigation
Documentation: Document all failures, investigations, and resolutions in the Integration Verification Report

Release Verification

Model Integrity Verification

All released models include SHA-256 checksums for integrity verification:

{
  "models": {
    "icd_distribution_v1.1.0.pt": "sha256:abc123...",
    "erythema_intensity_v1.1.0.pt": "sha256:def456...",
    ...
  }
}

Validation Status

All models in this release have been validated according to R-TF-028-005 AI Development Report and meet or exceed their predefined performance criteria with 95% confidence intervals.

Known Limitations

ICD Category Distribution: Performance degrades for Fitzpatrick V-VI (meets threshold but with wider CI)
Dermoscopic images: Some models optimized for clinical images only

Release Checklist

This checklist confirms that all release deliverables have been prepared and verified by the AI Team in accordance with GP-028 procedure:

Note: Integration activities to be performed by the Software Development team are specified in GP-028 procedure, Section "Integration Verification Package"

Technical Support

The AI Team provides ongoing support to the Software Development team throughout the integration process, as specified in GP-028.

Support Channels

Support Type	Description	Response Time
Integration Queries	Questions about preprocessing, post-processing, or model configuration	24 business hours
Verification Failures	Assistance investigating integration verification discrepancies	8 business hours
Bug Reports	Model behavior issues identified during integration	4 business hours
Urgent Issues	Critical blockers preventing release	Immediate escalation

Contact Information

Primary Contact: AI Team Lead
Communication Channel: Dedicated Slack channel #ai-integration-support
Issue Tracking: JIRA project AI-INT

Support Scope

The AI Team provides support for:

Technical Clarifications: Explaining preprocessing requirements, output formats, and configuration parameters
Verification Assistance: Helping investigate and resolve integration verification failures
Configuration Optimization: Advising on optimal runtime configurations for specific deployment environments
Documentation Updates: Addressing gaps or ambiguities in release documentation

Escalation Path

Level 1: AI Team Engineer (routine queries)
    ↓
Level 2: AI Team Lead (complex issues, verification failures)
    ↓
Level 3: Head of AI Development (critical blockers, release decisions)

Signature meaning

The signatures for the approval process of this document can be found in the verified commits at the repository for the QMS. As a reference, the team members who are expected to participate in this document and their roles in the approval process, as defined in Annex I Responsibility Matrix of the GP-001, are:

Author: Team members involved
Reviewer: JD-003 Design & Development Manager, JD-004 Quality Manager & PRRC
Approver: JD-001 General Manager

ㅤㅤ ㅤㅤ

Purpose​

Scope​

Related Documents​

Package Details​

Overview​

Model File Format​

Configuration Files​

Clinical Models: Detailed Specifications​

ICD Category Distribution and Binary Indicators​

ICD Category Distribution Algorithm​

Purpose​

Input Specifications​

Output Specifications​

Preprocessing Requirements​

Post-processing Requirements​

Configuration File​

Performance Summary​

Expected Results​

Binary Indicators Algorithm​

Purpose​

Input Specifications​

Output Specifications​

Calculation Method​

Configuration File​

Performance Summary​

Expected Results​

Visual Sign Intensity Quantification Models​

Common Specifications​

Input Specifications (All Intensity Models)​

Output Specifications (All Intensity Models)​

Preprocessing (All Intensity Models)​

Post-processing (All Intensity Models)​

Architecture​

Erythema Intensity Quantification​

Desquamation Intensity Quantification​

Induration Intensity Quantification​

Pustule Intensity Quantification​

Crusting Intensity Quantification​

Xerosis Intensity Quantification​

Swelling Intensity Quantification​

Oozing Intensity Quantification​

Excoriation Intensity Quantification​

Lichenification Intensity Quantification​

Wound Characteristic Assessment Models​

Input Specifications​

Output Specifications​

Lesion Quantification Models (Object Detection)​

Common Object Detection Specifications​

Input Specifications​

Output Structure​

Post-processing​

Inflammatory Nodular Lesion Quantification​

Acneiform Lesion Type Quantification​

Hive Lesion Quantification​

Surface Area Quantification Models (Segmentation)​

Common Segmentation Specifications​

Input Specifications​

Output Structure​

Post-processing​

Wound Surface Quantification​

Erythema Surface Quantification​

Hair Loss Surface Quantification​

Nail Lesion Surface Quantification​

Performance Summary​

Configuration File​

Expected Results​

Hypopigmentation/Depigmentation Surface Quantification​

Hyperpigmentation Surface Quantification​

Pattern Identification Models​

Follicular and Inflammatory Pattern Identification​

Input Specifications​

Output Specifications​

Performance Summary​

Configuration File​

Expected Results​

Inflammatory Pattern Identification (Hurley Staging)​

Output Specifications​

Hair Follicle Quantification​

Performance Summary​

Configuration File​

Purpose

Scope

Related Documents

Package Details

Overview

Model File Format

Configuration Files

Clinical Models: Detailed Specifications

ICD Category Distribution and Binary Indicators

ICD Category Distribution Algorithm

Purpose

Input Specifications

Output Specifications

Preprocessing Requirements

Post-processing Requirements

Configuration File

Performance Summary

Expected Results

Binary Indicators Algorithm

Purpose

Input Specifications

Output Specifications

Calculation Method

Configuration File

Performance Summary

Expected Results

Visual Sign Intensity Quantification Models

Common Specifications

Input Specifications (All Intensity Models)

Output Specifications (All Intensity Models)

Preprocessing (All Intensity Models)

Post-processing (All Intensity Models)

Architecture

Erythema Intensity Quantification

Desquamation Intensity Quantification

Induration Intensity Quantification

Pustule Intensity Quantification

Crusting Intensity Quantification

Xerosis Intensity Quantification

Swelling Intensity Quantification

Oozing Intensity Quantification

Excoriation Intensity Quantification

Lichenification Intensity Quantification

Wound Characteristic Assessment Models

Input Specifications

Output Specifications

Lesion Quantification Models (Object Detection)

Common Object Detection Specifications

Input Specifications

Output Structure

Post-processing

Inflammatory Nodular Lesion Quantification

Acneiform Lesion Type Quantification

Hive Lesion Quantification

Surface Area Quantification Models (Segmentation)

Common Segmentation Specifications

Input Specifications

Output Structure

Post-processing

Wound Surface Quantification

Erythema Surface Quantification

Hair Loss Surface Quantification

Nail Lesion Surface Quantification

Performance Summary

Configuration File

Expected Results

Hypopigmentation/Depigmentation Surface Quantification

Hyperpigmentation Surface Quantification

Pattern Identification Models

Follicular and Inflammatory Pattern Identification

Input Specifications

Output Specifications

Performance Summary

Configuration File

Expected Results

Inflammatory Pattern Identification (Hurley Staging)

Output Specifications

Hair Follicle Quantification

Performance Summary

Configuration File