Skin Lesion Classifier

Live Demo: https://ham10000-skin-lesion-classification.streamlit.app/

Dataset

• This project uses the HAM10000 dataset containing 10,015 dermatoscopic images across 7 skin lesion classes. Data was split into 70% training, 15% validation, and 15% testing.

• AI-powered skin lesion classification using deep learning, achieving 84% recall on melanoma detection.

Features

• ResNet18-based classifier trained on HAM10000 dataset
• Grad-CAM visualization for model interpretability
• 7-class classification (melanoma, nevus, basal cell carcinoma, etc.)
• 75% overall accuracy, 84% melanoma recall

Tech Stack

• Python, PyTorch, Torchvision, Streamlit, OpenCV, NumPy, Scikit-learn, Matplotlib, PIL

Technical Details

Training

• Model - ResNet18(Trained On ImageNet)

• Loss - CrossEntropyLoss

• Optimizer - AdamW

• Epochs - 25

• Learning Rate - 0.003 to 0.001

          precision    recall  f1-score   support
  
      nv       0.97      0.72      0.83      1006
     mel       0.33      0.86      0.48       167
     bkl       0.66      0.65      0.65       165
     bcc       0.71      0.83      0.77        77
   akiec       0.64      0.37      0.47        49
    vasc       0.73      0.86      0.79        22
      df       0.58      0.65      0.61        17
  

APP Demo

Grad_CAM Visualization

Installation

• git clone https://github.com/vnithin9632-ux/ham10000-skin-lesion-classification.git
• cd project
• pip install -r requirements.txt

For Running The App Locally

• streamlit run app.py

Usage

• Upload a skin lesion image to get instant classification with confidence scores and attention heatmap.

Limitations

• "Performance may drop on non-dermoscopic images"
• "Dataset imbalance affects minority classes"
• "Not intended for clinical diagnosis"

⚠️ This tool is for educational purposes only and not for medical diagnosis.