🔬 DermaScan AI — Skin Cancer Detection

A complete skin lesion segmentation + classification pipeline using:

• U-Net (with Attention Gates) for lesion masking
• MobileNetV2 for binary classification (Benign / Malignant)
• Streamlit for the interactive web UI

---

📁 Project Structure

skin_cancer_project/
├── app.py                    # Streamlit web application
├── train.py                  # Training script (run in Colab)
├── requirements.txt
├── models/
│   ├── __init__.py
│   ├── unet.py               # Refined U-Net with attention gates
│   └── classifier.py         # MobileNetV2 classifier head
└── utils/
    ├── __init__.py
    ├── datasets.py            # Dataset classes & transforms
    ├── training.py            # Trainers, losses (Dice + BCE), metrics
    └── inference.py           # End-to-end inference pipeline

---

🚀 Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Run the Streamlit app

streamlit run app.py

Set the model paths in the sidebar to point to your trained .pth files.

---

🏋️ Training (Google Colab)

Expected directory layout on Google Drive:

/content/drive/MyDrive/SkinCancerProject/
├── data/
│   ├── images/          ← raw ISIC dermoscopy images (.jpg)
│   ├── masks/           ← binary segmentation masks (.png)
│   │                      naming: <image_id>_segmentation.png
│   ├── cropped/         ← lesion crops for classification
│   └── labels.csv       ← two-column CSV: image_id, label (benign/malignant)
└── models/
    ├── unet_best.pth     ← saved by train.py
    └── mobilenet_best.pth

Run training in Colab:

# Mount drive first, then:
!python train.py --mode both    # train both models
!python train.py --mode seg     # segmentation only
!python train.py --mode cls     # classification only

---

---

📊 Inference Pipeline

Input Image
    │
    ▼
[U-Net]  →  Probability mask (0–1)
    │
    ▼
Otsu Threshold + Morphological cleanup  →  Binary mask
    │
    ├── Overlay on original image  →  Display
    │
    ▼
Crop lesion (bounding box + padding)
    │
    ▼
[MobileNetV2]  →  P(Malignant)
    │
    ▼
Label: Benign / Malignant + Confidence

---

⚕️ Disclaimer

This tool is for research and educational purposes only. It does not constitute medical advice. Always consult a qualified dermatologist.