ResearchJul 2025 to May 2026
Project Phoenix
Catch the earliest cell changes in cervical cancer, and show why the model decided.
- Convolutional neural networks
- Explainable AI
- Medical imaging
- Python
What I did
- Enhanced contrast and reduced noise in the SipakMed and Herlev cell images, keeping the features a diagnosis depends on.
- Trained convolutional neural networks to sort cells into five types.
- Used visual explanations to highlight the regions driving each prediction, so a doctor can check the model's reasoning.
- Put the model in the browser, at phoenix.meetbhatt.com.
The aim
Cervical cancer can be caught early from images of cells, but a model a doctor can't question is hard to trust. Phoenix sorts each cell into one of five types, and shows which parts of the image it based that on.