MedNIST = 64x64 grayscale; uploads are downsampled - the model has not seen full-resolution clinical images. Predictions on real clinical imagery will mislead. The 'Known failure modes' panel below is the lesson.
Success cases (in-distribution). The model should classify each of these MedNIST tiles correctly. GradCAM heatmap should land on the relevant anatomy. Click one to load it, then tick the confirmation and hit Run.
Failure cases (augmentation-induced). Each tile below is a MedNIST image that has been perturbed in a clinically plausible way (rotated film, inverted window, narrow contrast, noisy acquisition). The model classifies these confidently and wrong. Click one and run the demo to see GradCAM heatmap drift off the anatomy. See the preview gallery below for what each result will look like.
Preview: what the 4 failure cases look like
Each composite below shows the augmented input on the left and the GradCAM heatmap on the right. The heatmap visibly drifts off the anatomy onto edges, noise, or background gradients - the model is confidently wrong AND looking at the wrong region. Click any failure tile in the input panel above to reproduce these results live.