DermAI Clinical Screen is a high-sensitivity computer vision model optimized for the classification of dermatoscopic skin lesions. It is the primary inference engine for the DermAI Clinical Screening Space.

📝 Model Details

Architecture: ConvNeXt-Tiny (Hybrid Convolutional-Transformer)
Task: 7-class Multi-class Classification
Classes: AKIEC, BCC, BKL, DF, MEL (Melanoma), NV, VASC
Primary Objective: High recall for malignant lesions (Melanoma/BCC) to support clinical triage.
Framework: PyTorch & Hugging Face Hub

🎯 Intended Use

Primary Use: Decision support for dermatologists and clinical practitioners.
Input: High-quality dermatoscopic images (magnified skin surface).
Output: Probability distribution across 7 diagnostic categories + Triage recommendation.

🔬 Clinical Methodology & Features

1. Uncertainty Estimation (MC Dropout)

This model is configured to support Monte Carlo Dropout during inference. By performing 10+ forward passes with dropout enabled, the model calculates a Stability Score, helping to identify cases where the AI is "unsure" due to out-of-distribution features.

2. Explainability (XAI)

The model is compatible with Grad-CAM, allowing clinicians to visualize the specific spatial features (e.g., irregular pigment networks or blue-white veil) that led to the AI's classification.

3. ABCDE Feature Extraction

When integrated with the DermAI frontend, the model's predictions are augmented with automated geometric analysis:

A (Asymmetry): Calculation of symmetry across orthogonal axes.
B (Border): Compactness and irregularity metrics.
C (Color): Variance in the RGB/HSV color space.
D (Diameter): Pixel-to-mm estimation for lesion size.

📊 Training Data & Performance

Dataset: HAM10000 (10,015 images).
Optimization: Fine-tuned from ImageNet-1K weights with Focal Loss to handle class imbalance.
Preprocessing: 224x224 input resolution, normalization, and heavy data augmentation to simulate varied lighting conditions.

⚠️ Limitations & Disclaimers

Not for Autonomous Diagnosis: This model is a tool for clinicians and should not be used as a standalone diagnostic device.
Hardware Requirement: Best performance achieved on dermatoscopic images; standard mobile camera photos (non-dermatoscopic) may lead to reduced accuracy.
Bias: Performance may vary based on skin tone representation within the training dataset.

🛠️ How to Load (Python)

python

from huggingface_hub import hf_hub_download
import torch

# 1. Download the weights
weights_path = hf_hub_download(
    repo_id="imtiazhumzah/DermAI-Clinical-Screen", 
    filename="best_melanoma_recall_model.pth"
)

# 2. Load onto your model (assuming 'model' is already initialized)
state_dict = torch.load(weights_path, map_location='cpu')
model.load_state_dict(state_dict)
model.eval()

print("DermAI weights loaded successfully!")