Skin Cancer Classification using Transfer Learning and Handcrafted CNN

This repository contains the implementation of a skin cancer binary classification project using transfer learning models and a handcrafted CNN. The models are designed to classify images as either benign or malignant with high accuracy. Additionally, the project includes a user-friendly web application that allows users to upload skin lesion images and select a model for predictions, providing a confidence percentage for each result.

Key Features:

• Pre-Trained Models: Implementations of ResNet50, EfficientNetB5, MobileNetV2, and VGG16, trained on the ISIC dataset for binary classification.
• Handcrafted CNN: A custom-designed convolutional neural network inspired by VGG16 architecture, achieving an accuracy of 87%.
• Ensemble Approach: Combines predictions from all models to improve overall accuracy and robustness.
• Web Interface: A simple and interactive web application for model selection and image classification, making AI-assisted diagnostics accessible.
• Visualization Tools: Confusion matrices, ROC curves, and loss/accuracy graphs to evaluate model performance.

Getting Started:

The repository includes all necessary scripts for model training, evaluation, and deployment.

Dataset:

skin-cancer-isic-images

Pretrained Models Weights

Find weights here

Install Requirements

pip install flask flask-cors torch torchvision

Run server

python server.py

Run Frontend

Open index.html in any browser

Skin Cancer Detector Demo

Detector.Demo.mp4

Colabrators

• Dina Ashraf
• Iman Mohamed
• Nadine EL-Qersh

License

MIT License