Grad-CAM: Where a Medical-Scan AI Is Looking (2D)

Synthetic Lesion

Horizontal position 50%

Vertical position 50%

Lesion contrast 40%

Explain Class

CAM Source Layer

Classifier Output

P(Abnormal)

—

P(Normal)

—

Localization IoU vs true lesion

—

Peak raw CAM activation

—

How it works (2D)

Same real math as the 3D version — an actual small CNN (fixed, reproducible weights) processes a synthetic 32×32 scan: two real convolutional layers (edge/blob/blur filters → ReLU → 2×2 max-pool), a global-average-pool, and a linear head scoring "Normal" vs "Abnormal" — laid out here as flat 2D heatmap panels instead of 3D bars.

Class Activation Map (Zhou et al. 2016;
generalized as Grad-CAM by Selvaraju et al. 2017):

  L^c(i,j) = ReLU( Σ_k w_k^c · A_k(i,j) )

  A_k(i,j)  = k-th feature-map channel, final conv layer
  w_k^c     = ∂y^c / ∂(GAP(A_k))
            = classifier weight for channel k, class c
            (exact when the head is GAP + linear, as here)

Fixed vs the 3D source: the 3D sim wires its lesion-selective deep channel to the Laplacian/blob filter. This scan's background texture has a high-frequency term whose curvature is larger than the smooth lesion's, so after ReLU that channel saturates on background noise — verified numerically, P(Abnormal) stayed pinned at 99.94% across the whole 0–60% contrast range in the 3D sim. This 2D version wires the lesion-selective channel to the blur/local-brightness filter instead — smoothing suppresses the high-frequency background far more than it suppresses a smooth Gaussian bump, so the classifier genuinely responds: P(Abnormal) now swings from ≈12% (no lesion) to ≈85%+ (strong, centered lesion), and the CAM heatmap peak lands within one grid cell of the true lesion center.

Position / contrast sliders — move a real Gaussian-bump "lesion" through the image and control how strongly it stands out from the background texture.
Explain Class — chooses which class's score y^c is being explained; the weight vector w^c flips, so the same feature maps produce a different heatmap.
CAM Source Layer — Final conv shows the true gradient-weighted CAM (8×8, coarse but class-discriminative); Layer 1 blur shows the raw pre-classifier response (32×32, sharp but not class-aware).
Panel layout — input scan (top-left) → CAM overlay (bottom-left) → Layer-1 feature maps (4 tiles) → Layer-2 feature maps (6 tiles) → output probability bars (bottom).

Real-world relevance: this is the exact technique radiologists and regulators demand from black-box scan classifiers before trusting a "positive" call — the heatmap must land on the actual lesion, not on an unrelated artifact, and the predicted class must actually track the pathology, not the background.

Low activation Mid activation High activation