SDLC Corp

🧠 Brain Tumor Detection

Two models run entirely in your browser — no server, no upload: a ViT predicts the tumor class (Glioma / Meningioma / No-tumor / Pituitary), and a ResNet50 produces an instant class-activation heatmap.

⚠️ For educational and research use only. Not a medical device. Do not use for diagnosis. Always consult a qualified radiologist.

1 · Choose an MRI scan

Click to upload · or drag an image here

Or try an example:

2 · Result

Loading model…

Test-time augmentation

ℹ️ About & how it works

An SDLC Corp application. Two models are exported to ONNX (float16) and run client-side with onnxruntime-web — your image never leaves your device:

A Vision Transformer (ViT) produces the prediction, confidence and uncertainty (≈97% on a held-out test set — see README).
A ResNet50 produces the grad-free class-activation heatmap and the feature embedding used for OOD detection and similar-scan retrieval.

The CAM, test-time augmentation, uncertainty, OOD and retrieval pipeline are SDLC Corp's own work.

Class Activation Map (CAM): because the network ends in a global-average-pool followed by a single linear layer, the heatmap can be computed without gradients as a weighted sum of the last convolutional feature maps (CAM_c = Σ_k W_c,k·A_k). It's instant — and you can click any class to see where the model looks for that class. Bright regions contribute most to the score.

Advanced ML, all in-browser:

Test-time augmentation — averages predictions over flipped views (self-ensembling) for a more robust result.
Uncertainty — predictive entropy (+ cross-view agreement) gives a certainty meter and flags low-confidence cases.
Out-of-distribution detection — compares the image's feature embedding to known MRIs; warns if it doesn't resemble any (e.g. a non-MRI upload).
Similar-scan retrieval — pools the conv features into a 2048-d embedding and finds the nearest reference scans by cosine similarity (content-based image retrieval).