XRayLens
A pneumonia classifier gives you an answer. This lab shows how it got there, how far you can trust it, and where it breaks.
Not for diagnosis
Choose a radiograph
Start with the curated contact sheet. Each case was picked from the evaluation to teach one thing. You can also upload a public, de-identified research image. It is analysed on your device and never uploaded anywhere.
Use only public, de-identified research images. Do not upload real patient records. DICOM tags other than modality and view are ignored.
Read the model’s output
The model returns two softmax scores that sum to 1. A fixed threshold of 0.5 turns them into a label.
See where the model looked
Grad-CAM weights the last convolutional block’s feature maps by how strongly they push the chosen class’s logit up. Bright regions influenced the output. That does not make them correct.
×1.00 is what a uniform, uninformative map would score. Heat on padding or the frame (text markers, tubes, collimator edges) is a shortcut warning.
Sanity check: does this map depend on what the model learned?
Stress-test this image
Eight edits that do not change the anatomy a radiologist would read. A trustworthy model’s score should hardly move.