confusion matrix on the test set: nearly all scans fall on the diagonal for glioma, meningioma, no tumor and pituitary
why
a basic CNN trained from scratch needed augmentation to generalize, but natural-photo-strength augmentation distorted the exact regions MRI diagnosis depends on.
how it works
four conv/pool blocks, mild rotation/zoom/brightness augmentation tuned specifically for MRI, trained with early stopping and checkpointing.
what i learned
the first augmentation pass dropped accuracy to 92.5%. turning the strength down to roughly a third fixed it and beat the no-augmentation baseline on every class.
results
97.15% test accuracy, 94.3% meningioma F1, documented end to end in the README.
limitations are documented in the README, not claiming clinical validity.
