Brain Tumour Atlas

Reading the brain,
one specimen
at a time.

An interactive atlas. Upload a brain MRI and a deep neural network classifies it across four diagnostic categories: glioma, meningioma, pituitary, or none. It also shows you where in the scan it looked.

Begin a reading How it works

Editorial atlas plate -- axial section of a cerebral tumor

Plate XXVII -- tumor cerebriaxial section

From the editorial atlas, plate XXVII. Click upload to load your own.

Console

Run a reading.

Specimen

Bring a scan.

Drag & drop an MRI here

JPG, PNG, WebP. Max 10 MB.

Reading

The verdict.

No reading yet.

Choose a specimen above, pick a model, and click classify. The prediction, attention overlay, and a written interpretation will appear here.

Atlas

Four scenarios.

Glioma
Plate 01
Tumors arising from glial cells in the brain or spine.
Meningioma
Plate 02
Slow-growing tumors of the meninges membrane.
Pituitary
Plate 03
Abnormal growths within the pituitary gland.
No Tumor
Plate 04
Healthy scan classification baseline.

Method

From MRI to analysis.

Pipeline

Three steps, in order

Preprocess

Decode the upload, resize to the model's input resolution (299x299 for Xception, 224x224 for the custom CNN), and normalise pixel values to [0, 1].

Infer

Forward pass through the chosen model. The network produces a softmax over four classes; we take the argmax as the prediction and keep the full distribution for display.

Interpret

TensorFlow's GradientTape gives a per-pixel saliency map of the predicted logit. Gemini reads the overlay back to you in plain English, in four sentences.

Appendix

Click to open

7,023 MRIs from the Kaggle Brain Tumor MRI dataset.
5,712 train, 655 validation, 656 test. The test set is split 50/50 from the original Testing folder, stratified by class with a fixed random seed.
Xception: ImageNet-pretrained backbone with a small dense classifier on top. Trained for 5 epochs at 299x299.
ResNet50V2: ImageNet-pretrained backbone using the same classifier head as Xception, for a side-by-side architecture comparison at 299x299.
Custom CNN: four conv-pool blocks (512, 256, 128, 64 filters) and a 256-unit dense head. Trained for 10 epochs at 224x224.
All three use the Adamax optimiser at learning rate 0.001.
Test-set accuracy is high (>99% on all three models). During interpretability checks the saliency overlay sometimes attends to cortical surface features even when predicting tumours that anatomically live elsewhere. For example, pituitary lesions live at the sella turcica, not the cortex. Two likely causes: vanilla gradient saliency is noisy and biased toward high-frequency edges, and per-class folders in the dataset may carry distinct preprocessing artefacts that the network has learned as shortcuts (a Clever Hans signal).
Useful next steps would be Grad-CAM or Integrated Gradients for better localisation, plus per-folder preprocessing audits to rule out the shortcut hypothesis. The takeaway for now: the classifier is accurate, but the attention maps should not be read as a radiologist's reasoning.

Reading the brain,one specimenat a time.

Glioma

Meningioma

Pituitary

No Tumor

Preprocess

Infer

Interpret

Reading the brain,
one specimen
at a time.