Brain-Tumour-MRI---Segmentation-and-Classification

Brain Tumor MRI Segmentation and Classification

A multi-task model that looks at a brain MRI scan and does two jobs at once. It draws the tumor region (segmentation) and it names the scan (classification) as one of four classes: glioma, meningioma, pituitary, or no_tumor. Two architectures are trained and compared, a standard U-Net and an Attention U-Net, both in PyTorch. The dataset is BRISC2025.

The notebook in notebooks/ is the original project, kept with all of its outputs. The src/ folder holds the same pipeline as Python modules.

The problem

Segmentation answers "where is the tumor" by labelling every pixel. Classification answers "what kind of scan is this". A single network handles both: a U-Net encoder and decoder produce the mask, and a small head on the bottleneck features produces the class. The two losses are combined as 0.7 times Dice for the mask plus 0.3 times cross-entropy for the class.

Data

BRISC2025 ships a segmentation task and a classification task. The segmentation task holds tumor scans paired with masks. The classification task holds scans in four class folders, including healthy scans under no_tumor. Images are grayscale and are resized to 256 by 256, denoised with a light Gaussian blur, and the masks are cleaned with small morphological open and close steps.