Grad-CAM: Where a Medical-Scan AI Is Looking

Deep classifiers reading X-rays, CT slices and dermoscopy images are only trustworthy if a clinician can see why a model called a scan abnormal. This simulator runs a real, hand-specified two-layer convolutional network — actual 3×3 convolutions, ReLU, max-pooling, global average pooling and a linear head — over a synthetic scan containing a movable, adjustable-contrast lesion, then computes a genuine Class Activation Map: the gradient-weighted sum of the final convolutional layer's feature maps that reduces, for a global-average-pooled linear head, to the exact formula used in real Grad-CAM interpretability tooling. Drag the lesion around, watch the 3D bar-chart feature maps at each layer light up, switch which class is being explained, and compare the coarse-but-class-aware final-layer heatmap against a sharp-but-class-blind raw filter response from layer one — then check the live Intersection-over-Union score against the lesion's true location to see how faithfully the model's attention actually tracks the pathology.