02 / Computer vision · Deep learning

Brain Tumor Detection

Interpretable MRI classification

An experimental pipeline that learns compact anatomical features with an autoencoder, classifies MRI scans from the latent representation and uses Grad-CAM to inspect the regions influencing each prediction.

TensorFlow / Keras
Autoencoder
Grad-CAM
OpenCV
scikit-learn

Type: Personal computer vision project
Dataset: Br35H — 450 MRI scans, 70/15/15 split
Outcome: 96% test accuracy · 450 images
Code: Full pipeline on GitHub ↗

Results from a real run

4 visuals

Grad-CAM heatmaps overlaid on brain MRI scans — Grad-CAM — the regions that drove each prediction

Autoencoder reconstruction samples — Autoencoder reconstructions — what the latent space keeps

Classifier training and validation metrics over epochs — Classifier training and validation metrics

Confusion matrix on the test set — Confusion matrix on the held-out test set

How it works

Approach

Learn the anatomy

A convolutional autoencoder compresses each 128x128 scan into a compact latent representation, forcing it to keep the patterns that matter.

Classify from the latent space

The encoder half feeds a binary classifier with dropout and batch normalization, evaluated with precision, recall and F1 on a held-out test set.

Explain every call

Grad-CAM heatmaps over the encoder's last convolutional layer show which regions pushed each decision — tumor and no-tumor cases alike.

See the proof
behind the case.

GitHub repo ↗LinkedIn profile ↗

← Back to selected work