An end-to-end deep learning project that detects and classifies seven types of common skin lesions using the HAM10000 dataset and a CNN-based model. The application also provides medical insights and treatment suggestions based on the prediction.

Skin diseases, especially malignant ones like melanoma, can be life-threatening if not detected early. This project leverages the power of Convolutional Neural Networks (CNNs) and Transfer Learning to classify dermatological images and guide users with basic treatment advice.

• Classifies skin lesions into 7 categories
• Uses EfficientNet/ResNet for high accuracy
• Web interface for uploading images
• Provides confidence scores, medical names, and first-aid/treatment suggestions
• Optional Grad-CAM visualization to highlight lesion areas

Name: HAM10000 - Human Against Machine with 10000 training images
Source: Kaggle

We used Transfer Learning with fine-tuned CNN models such as:

• EfficientNetB0
• ResNet50

• Image Input Layer (224x224)
• Convolutional Base from pretrained model
• Global Average Pooling
• Dense + Dropout Layers
• Softmax Output (7 classes)

• Categorical Crossentropy
• Accuracy, Precision, Recall, F1-score

• Resizing images to 224x224
• Normalizing pixel values (0–1)
• One-hot encoding labels
• Train-Validation-Test Split (70/15/15)
• Class balancing using data augmentation

• Accuracy: ~90+% (Varies with model)
• Confusion matrix and classification report used for performance analysis

• Upload skin image
• Get:
  • Predicted condition
  • Confidence level
  • Medical name and short description
  • First-aid and treatment guidance
• View highlighted region (optional Grad-CAM)

git clone https://github.com/your-username/skin-disease-classifier.git
cd skin-disease-classifier
pip install -r requirements.txt
streamlit run app.py

Derma-AI/
├── data/
│   ├── HAM10000_images/
│   └── metadata.csv
├── model/
│   └── skin_disease_model.h5
├── app.py
├── utils/
│   ├── preprocess.py
│   ├── gradcam.py
├── treatment_info.json
├── requirements.txt
└── README.md

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