Deep Learning · MRI · Interpretability

Reading the
brain, honestly.

A transfer-learning system that detects and classifies brain tumors from MRI, and shows you exactly where it looked to decide.

⚠ Research & education only. Not a medical device.
3D MODEL · drag to rotate
glioma zone meningioma occipital pituitary / base cerebellum
Overview

Two models, one honest pipeline

Built on EfficientNetB0 with ImageNet weights, fine-tuned in two stages, evaluated on clean stratified splits with no data leakage. Real numbers.

99.4%
Binary accuracy
94.75%
4-class accuracy
±0.40
Std over 3 seeds
0.987
No-tumor recall

Detection

Binary tumor vs. no-tumor from a merged dataset, handling class imbalance with damped weighting.

recall (no-tumor): 0.987

Typing

Four classes (glioma, meningioma, pituitary, no-tumor) trained with label smoothing and staged fine-tuning.

macro-F1: 0.950

Interpretability

Grad-CAM overlays confirm the model attends to real lesions, not skull or scanner artifacts.

method: Grad-CAM
Architecture

How a scan becomes a diagnosis

Hover any layer to see what it does. The scan flows left to right through the network to two prediction heads.

EfficientNetB0

Hover a layer above to inspect it. The pretrained backbone does most of the visual work; the two lightweight heads specialize it for detection and typing.

Results

Per-class performance

Multiclass F1 across the four categories on the held-out test set.

glioma 0.942
meningioma 0.912
no tumor 0.965
pituitary 0.948

Multiclass confusion matrix

Rows = true, columns = predicted. Meningioma is the hardest class, occasionally confused with glioma.

The honest story

The original project reported 78.7% multiclass. Diagnosis revealed the cause wasn't the model. It was tangled preprocessing and premature early stopping. After fixing the pipeline and evaluating on clean stratified splits:

before: 78.7%  →  after: 94.75% ± 0.40%
verified over 3 independent random seeds.
Interpretability

Where the model looks

Grad-CAM overlays the model's attention on the scan. Tight focus on the lesion means the prediction rests on real anatomy, and the failures explain themselves.

Grad-CAM across four classes
Grad-CAM across all four classes. Heat concentrates on each tumor type.
Clinical viewer meningioma
Clinical-viewer style: grayscale MRI beside the heat overlay + impression.
Demo

The interactive interface

Upload a scan, get both predictions, see the Grad-CAM, and ask the built-in assistant why. Every answer is grounded in the model's own output.

NeuroScan interface
Glioma case: detection, 4-class typing, Grad-CAM focus, and grounded chat.
View the code on GitHub →
Findings

What we learned