A web-based system to detect skin diseases from images using deep learning. The project uses ensemble learning with CNN architectures to improve diagnostic accuracy.

Users can upload a skin image, and the backend uses three trained CNNs (EfficientNetB3, ResNet101, DenseNet121) to classify it. The predictions are averaged to determine the final result.

• 🖼️ Upload dermatoscopic images for prediction
• 🧠 Ensemble of EfficientNetB3, ResNet101, and DenseNet121
• ✅ JWT-based authentication system (register/login)
• 📊 Confidence score for each prediction
• 📖 Description of predicted skin condition
• 📱 Responsive UI using React 18

• Python 3.12
• Flask + Flask-JWT-Extended
• TensorFlow / Keras
• MongoDB (via PyMongo)
• OpenCV & Albumentations for preprocessing
• dotenv for environment configs

• React 18
• Axios
• React Router DOM
• Styled Components
• React Icons

skin_disease_detection/
├── backend/
│ ├── app/ # Backend app logic
│ ├── main.py # Flask app entry
│ ├── requirements.txt # Python dependencies
│ └── uploads/ # Uploaded images
├── frontend/
│ └── src/ # React app
├── trained_models/ # Pretrained .h5 models
└── README.md

• Dataset: HAM10000 (Kaggle)
• 10,000 labeled images across 7 classes:
  • akiec, bcc, bkl, df, mel, nv, vasc

To prepare:

• Extract dataset into:
  backend/data/skin_disease_dataset/base_dir/
  with subdirectories: train_dir/, val_dir/, test_dir/

• Ensemble of:
  • ✅ DenseNet121
  • ✅ EfficientNetB3
  • ✅ ResNet101
• Each model is trained independently.
• Final prediction: average of softmax scores from all 3 models.

Download the pre-trained models from Google Drive:

🔗 Download Models

After downloading, extract and place the trained_models folder in the project root directory:

skin_disease_detection/
├── backend/
├── frontend/
├── trained_models/      ← Place here
│   ├── densenet121.h5
│   ├── efficientnetb3.h5
│   └── resnet101.h5
└── README.md

• JWT-based login/register system
• Tokens must be passed in Authorization header for prediction requests.

• Python 3.12+
• Node.js v18+
• MongoDB (local or cloud)
• Git

# Navigate to backend
cd backend

# Create virtual environment
python -m venv venv

# Activate virtual environment
# On Windows:
venv\Scripts\activate
# On macOS/Linux:
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

FLASK_SECRET_KEY=YourSecretKey
JWT_SECRET_KEY=YourJWTSecretKey
MONGO_URI=mongodb://localhost:27017/skin_disease_db

If you want to train the models yourself instead of using the pre-trained ones, download and extract the HAM10000 dataset into:

backend/data/skin_disease_dataset/base_dir/
├── train_dir/
├── val_dir/
└── test_dir/

python main.py

cd frontend
npm install
npm start

Login → Receive JWT token

Upload skin image

Backend runs predictions using all 3 models

Softmax probabilities are averaged

Highest scoring class is selected

Response includes:

Predicted disease

Confidence score

Disease name + description

MIT License See LICENSE file.

@Purvesh-PJ

Detailed project documentation is available in the docs/ folder:

• Architecture Overview - System design and component breakdown
• Working Flow - End-to-end process flows
• Modules Documentation - Detailed module and component reference
• Frontend Flow - React architecture and patterns
• Backend Flow - Flask API and request lifecycle
• API Overview - Complete API endpoint reference
• Interview Notes - Interview preparation guide
• Improvements - Future enhancements and recommendations

• Kaggle HAM10000 Dataset
• TensorFlow
• React
• Flask