DermTrace AI — Skin Lesion Classification & Explainability System 🩺

Samsung Innovation Campus & EOI — AI Capstone Project
DermTrace AI is a deep learning system designed for multi-class skin lesion diagnosis (HAM10000 dataset), incorporating DullRazor hair artifact removal, two-phase transfer learning (ResNet152 / Xception), class imbalance correction, and Grad-CAM visual explainability.

---

📌 Key Features

• DullRazor Preprocessing: Mathematical morphology (Black-Hat filter + Telea inpainting) to eliminate hair artifacts without distorting underlying lesion pigment networks.
• Anti-Leakage Partitioning: Strict lesion_id grouping (GroupShuffleSplit) to prevent artificial metric inflation caused by duplicate images of the same lesion.
• Two-Phase Transfer Learning:
  • Phase 1 (Feature Extraction): Backbone frozen (ResNet152 / Xception), training ~1.06M head parameters with $\text{LR}=10^{-3}$.
  • Phase 2 (Fine-Tuning): Unfreezing deep convolutional blocks with low learning rate ($\text{LR}=10^{-5}$).
• Class Imbalance Mitigation: Dynamic real-time data augmentation + class-weighted categorical cross-entropy loss.
• Visual Explainability (Grad-CAM): Gradient-weighted Class Activation Mapping generating heatmaps that highlight key diagnostic regions driving network predictions.
• Streamlit Web Application: Interactive web interface featuring drag-and-drop file upload, live mobile camera capture, side-by-side preprocessed views, and risk severity alerts.

---

📊 Benchmark & Experimental Results (E1 – E8)

Exp ID	Model Architecture & Pipeline	Split Method	Balanced Acc (%)	Macro-F1 (%)	Melanoma Recall (%)
E1a	Baseline (MobileNetV2, no DullRazor)	Image-level	74.2%	63.5%	54.8%
E1b	MobileNetV2 + DullRazor Filter	Image-level	78.5%	68.2%	61.3%
E2	ResNet152 (Trained from scratch)	Lesion-level	45.1%	38.0%	29.4%
E3	ResNet152 (Without Class Weights)	Lesion-level	68.4%	58.1%	42.0%
E4	ResNet152 + DullRazor + Class W.	Lesion-level	92.4%	88.7%	89.5%
E5	Xception + DullRazor + Class W.	Lesion-level	91.8%	87.9%	88.2%
E7	ResNet152 (No Data Augmentation)	Lesion-level	81.3%	76.5%	71.0%
E8	ResNet152 (Image-level Leakage Test)	Image-level	97.8% (Inflated)	94.2% (Inflated)	95.1% (Inflated)

---

📁 Repository Structure

DermTrace_Clean_Repo/
├── .gitignore
├── README.md
├── requirements.txt
├── LICENSE
├── docs/
│   ├── Samsung_AI_Capstone_Project_Final_Report_DermTrace.md
│   └── Samsung_AI_Capstone_Project_Final_Report_DermTrace.docx
├── src/
│   ├── __init__.py
│   ├── dullrazor.py
│   ├── preprocess.py
│   └── gradcam.py
├── notebooks/
│   ├── 01_EDA_HAM10000.ipynb
│   └── 02_DermTrace_Colab_Training.ipynb
├── app/
│   ├── .streamlit/
│   │   └── config.toml
│   ├── app.py
│   └── labels.json
└── assets/
    └── screenshots/

---

🚀 Quickstart Guide

1. Local Environment Setup

# Clone repository
git clone https://github.com/username/DermTrace_Clean_Repo.git
cd DermTrace_Clean_Repo

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

2. Launch Streamlit Web Application

streamlit run app/app.py

---

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.