- Type
- Individual project
- Role
- Sole developer: design, development and testing.
- Tech stack
- Links
- Paper
Overview
A web application that takes a lung CT scan and classifies it as benign, malignant or normal, using a convolutional neural network designed for the job rather than a pre-trained model.
The custom network reached 98.55% accuracy, ahead of ResNet50 at 97.67% and VGG16 at 66.15% on the same data. Around the model sits a complete flow for the person using it: upload, analysis, results, guidance on what to do next, an appointment request by email, and a map of nearby specialists.
The work became the final year research project and was published at the ICTAR conference.
Built
- A custom CNN in TensorFlow, trained on 256x256 grayscale CT scans.
- A Flask web app for upload, analysis and results.
- Appointment requests sent with Flask-Mail.
- A Google Maps view of nearby specialists.
Learned
- Implementing TensorFlow for medical image processing.
- Training a model and deploying it behind a web application.
- Integrating outside services such as email and Google Maps.