This project focuses on the classification of skin lesions into seven categories using deep learning techniques. Leveraging the HAM10000 dataset (~10,000 dermatoscopic images), I developed a custom Convolutional Neural Network (CNN) and enhanced it further using transfer learning with ResNet50. Model explainability was incorporated through Grad-CAM to provide visual insights into predictions.

HAM10000 - Human Against Machine with 10000 training images
A curated collection of multi-source dermatoscopic images of common pigmented skin lesions:

• ~10,000 images
• 7 classes:
  • Melanocytic nevi (nv)
  • Melanoma (mel)
  • Benign keratosis-like lesions (bkl)
  • Basal cell carcinoma (bcc)
  • Actinic keratoses (akiec)
  • Vascular lesions (vasc)
  • Dermatofibroma (df)

• 3 Convolutional blocks with ReLU activation & MaxPooling
• Fully connected dense layers
• Batch Normalization & Dropout for regularization
• Softmax output layer for 7-class classification

• Pretrained on ImageNet
• Fine-tuned on HAM10000 data
• Custom top layers for adaptation to skin lesion classification

• Image augmentation: rotation, flipping, zoom, brightness
• Image normalization
• Dimensionality reduction for improved efficiency
• Early stopping & learning rate scheduler
• Grad-CAM for model explainability

Grad-CAM was used to generate class activation maps that highlight regions of the image contributing most to the model's prediction.

Example:

• Python
• TensorFlow / Keras
• Scikit-learn
• OpenCV
• Matplotlib & Seaborn
• Pandas & NumPy

├── skin_cancer_prediction.ipynb     # Main Jupyter Notebook (Kaggle)
├── data/                            # Contains HAM10000 images & metadata
├── models/                          # Saved model checkpoints
├── outputs/                         # Grad-CAM visualizations
└── README.md                        # Project overview