Brain Tumor MRI — CNN

Executive Summary

A computer-vision pipeline that classifies brain MRI scans into four categories (glioma, meningioma, pituitary, and no-tumor) using a custom convolutional neural network. The 3-block CNN (32 -> 64 -> 128 filters) uses Batch Normalization and Global Average Pooling and is trained on 5,700+ scans with targeted augmentation.

The Problem

Manual review of brain MRIs is slow and needs specialist radiologists. A reliable automatic classifier can triage scans and flag the likely tumor type for faster expert review — but medical-imaging models must generalise well and resist overfitting on limited data.

The Solution

• A custom 3-block convolutional network (32 -> 64 -> 128 filters) rather than a heavyweight pretrained backbone.
• Batch Normalization after convolutions for stable, faster training.
• Global Average Pooling instead of large dense layers to cut parameters and reduce overfitting.
• Targeted data augmentation (rotation up to 15 degrees) and plateau-based learning-rate decay.

System Architecture

1. Data & Augmentation Layer
   MRI scans resized to 224 x 224, with rotation augmentation (up to 15 degrees) to widen effective sample diversity.

2. Modeling Layer
   A 3-block CNN (Conv -> BatchNorm -> ReLU -> Pool) at 32 -> 64 -> 128 filters, followed by Global Average Pooling and a softmax head over 4 classes.

3. Training Layer
   Plateau learning-rate decay down to a 1e-6 floor, with feature-map visualisation to inspect what each block learns.

Engineering Decisions

Why Global Average Pooling?
GAP collapses each feature map to a single value, drastically reducing parameters versus flattened dense layers and acting as a structural regulariser against overfitting.

Why Batch Normalization?
It stabilises activations across the 3 blocks, allowing higher learning rates and faster convergence.

Pipeline Details

Dataset: 5,700+ MRI scans across 4 tumor categories.
Input: 224 x 224 resolution with rotation augmentation up to 15 degrees.
Architecture: 3-block CNN, 32 -> 64 -> 128 filters, BatchNorm + Global Average Pooling.
Training: plateau learning-rate decay to a 1e-6 floor.

Outcome

A parameter-efficient CNN image classifier for brain-tumor MRI triage, built with an emphasis on regularisation and interpretability over raw model size.