Grad-CAM

Grad-CAM computes the gradient of the score for a target class with respect to the feature maps of a convolutional layer. These gradients are globally average pooled to obtain neuron importance weights, which are then used to produce a weighted combination of the forward activation maps. The result is a coarse heatmap that highlights the discriminative regions the CNN used for its prediction. This approach is architecture-agnostic and works with any CNN without architectural changes or re-training.

Unlike earlier saliency methods that produce identical explanations regardless of the predicted class, Grad-CAM generates class-specific visualizations. By backpropagating the gradient from a specific output neuron, the resulting heatmap shows which image regions are relevant for that particular class. For example, in a medical image classified as 'pneumonia,' the heatmap will highlight the lung opacities, while for 'cardiomegaly' it will focus on the cardiac silhouette. This class-discriminative property is critical for multi-label diagnostic tasks.

Standard Grad-CAM produces low-resolution heatmaps due to the spatial downsampling in deep convolutional layers. Guided Grad-CAM fuses the coarse localization map from Grad-CAM with the fine-grained pixel-space gradients from Guided Backpropagation via pointwise multiplication. This hybrid approach yields visualizations that are both high-resolution and class-discriminative, showing precisely which pixels within the highlighted region contributed most to the decision. This is especially valuable in medical imaging where precise boundary delineation matters.

Grad-CAM can generate counterfactual explanations by negating the gradient of the target class score. Instead of highlighting regions that increase the class score, the resulting heatmap shows regions that would decrease the score if modified or removed. In a diagnostic context, this answers the question: 'What would need to change in this scan for the model to predict a different condition?' This capability supports clinician-in-the-loop workflows by showing what evidence would contradict the current diagnosis.

The faithfulness of Grad-CAM explanations can be measured using metrics like Drop in Confidence and Increase in Confidence. By progressively occluding the most salient regions identified by the heatmap and measuring the resulting change in the model's class score, one can quantify how accurately the explanation reflects the model's true reasoning. A faithful explanation will show a sharp monotonic drop in confidence as salient regions are removed. Tools like Quantus and Captum provide standardized implementations of these evaluation protocols.

Grad-CAM is widely used in FDA-cleared SaMD submissions and clinical validation studies to demonstrate that diagnostic models attend to clinically relevant anatomical regions rather than spurious correlations or artifacts. The technique supports regulatory explainability requirements by providing visual evidence that the model's decision basis aligns with established radiological criteria. When combined with domain-specific saliency constraints—such as anatomical atlases—Grad-CAM helps establish the trust calibration necessary for clinical adoption.

A gradient-free alternative to Grad-CAM that eliminates reliance on potentially noisy gradients. Instead, Score-CAM uses the increase in confidence on the target class as the weight for each activation map.

• Masks the input with each upsampled activation map and measures the forward-pass score change
• Avoids gradient saturation and vanishing gradient problems common in deep networks
• Produces less noisy localization maps, especially in deeper architectures

A computationally efficient attribution method that requires no gradient computation and no class-specific backpropagation. Eigen-CAM performs singular value decomposition on activation maps from convolutional layers.

• The first principal component of the activation tensor serves as the explanation
• Class-agnostic by default, making it useful for unsupervised anomaly detection
• Significantly faster than gradient-based methods, suitable for real-time clinical applications

A perturbation-based approach that measures the drop in class confidence when individual activation maps are zeroed out. Unlike Grad-CAM, it does not rely on gradient information at all.

• Computes importance weights through forward-pass ablation experiments
• Provides a direct causal measure: "How much does removing this feature map hurt prediction?"
• Computationally heavier than Grad-CAM but offers stronger faithfulness guarantees