Brain Tumor AI

The story

1. 01 · THE PROBLEM

   A label isn't evidence.

   A reviewer needs to know what the model predicted, how stable that answer is, and what part of the scan drove it.

   

2. 02 · THE MODEL

   A baseline, then something better.

   A scratch CNN is the control. EfficientNet-B0, fine-tuned from ImageNet with MC Dropout in its head, is the stronger model, reaching 89–90% accuracy (macro 0.90) across four classes on 7,200 MRIs. Pituitary and no-tumor recall is 97–99%; glioma and meningioma are harder at 0.73, which is exactly why every answer shows its evidence.

   

3. 03 · THE EVIDENCE

   Uncertainty and evidence, side by side.

   Monte Carlo Dropout keeps dropout active across 20 passes to expose instability. Grad-CAM paints where the network looked, so a wrong answer can be spotted rather than believed.

   

4. 04 · THE LESSON

   Medical AI credibility is visual and statistical.

   A classifier is easier to judge when uncertainty, comparison and visual evidence live in the same interface.

   

How it's built

Every box is a real part of the system. The dots are data moving through it.

The real thing

EfficientNet-B0 · accuracyEfficientNet-B0 · lossEfficientNet-B0 · confusionEfficientNet-B0 · ROCBaseline CNN · confusionBaseline CNN · ROC

Built with

• PyTorch
• EfficientNet-B0
• Grad-CAM
• Streamlit