OncoVision 🩺 - Skin Cancer Detection Model

OncoVision (from oncology + vision) is a project I built to help people check skin lesions for potential risks from home. It uses a CNN model to classify skin images as either Benign or Malignant with ~93% accuracy.

The goal isn't to replace a real doctor, but to help people catch potential issues early so they know when it's time to seek professional medical advice.

Demo

oncovision-demo.mp4

Features

• Classification: Categorizes skin lesion images using a deep learning model.
• Explainability: Generates Grad-CAM heatmaps to show exactly which parts of the image the AI is looking at.
• Fast UI: A clean, professional dashboard built with React and Vite.

Model Architecture:

• Convolutional layers with ReLU activation for feature extraction
• MaxPooling for spatial downsampling
• Dropout to prevent overfitting
• Fully connected layers leading to a sigmoid output for binary classification

Tech Stack

• Backend: Flask, TensorFlow, Keras, OpenCV
• Frontend: React, Vite, Framer Motion, Lucide Icons
• Model: Convolutional Neural Network (CNN) trained on thousands of labeled skin lesion images

Setup & Running Locally

The easiest way to run this is directly from source. You'll need two terminals open.

1. Backend (Flask)

cd server
pip install -r requirements.txt
python app.py

The server will run on http://localhost:5000.

2. Frontend (React)

cd client
npm install
npm run dev

The UI will be available at http://localhost:5173.

or if you have docker

docker build -t oncovision:latest .
docker run -d -p 80:80 -p 5000:5000 oncovision:latest

Dataset & Model

The model was trained on a dataset of skin images. You can find the raw data I used here.

Future Ideas

• Differentiating between specific cancer types (melanoma vs basal cell).
• Mobile optimization for easier photo taking.
• Improving the Grad-CAM resolution for better clinical insights.