Gradient-weighted Class Activation Mapping (Grad-CAM)

Grad-CAM computes the gradient of the score for a target class with respect to feature maps in the final convolutional layer. These gradients are globally average-pooled to obtain neuron importance weights, which are then used to produce a weighted combination of forward activation maps. The result is a coarse heatmap highlighting class-discriminative regions without architectural changes.

Unlike Class Activation Mapping (CAM), which requires a Global Average Pooling layer followed by a single fully-connected layer, Grad-CAM is applicable to any CNN architecture. It works with:

• VGG, ResNet, Inception, DenseNet
• Models with multiple fully-connected layers
• Networks with complex output heads This flexibility makes it the standard for post-hoc visual explanation in modern diagnostic imaging pipelines.

Standard Grad-CAM heatmaps are low-resolution due to the spatial downsampling in final convolutional layers. Guided Grad-CAM fuses the coarse localization map with pixel-space gradients obtained via guided backpropagation, producing fine-grained, class-discriminative visualizations. This is critical for medical applications where precise boundary delineation matters, such as tumor margin identification in radiology.

By negating the gradient of the target class score, Grad-CAM can generate counterfactual heatmaps that highlight regions which, if removed or altered, would increase the model's confidence in a different class. This is valuable for:

• Understanding why a model did not predict a specific condition
• Identifying confounding image features
• Auditing model reliance on spurious correlations in clinical datasets

Grad-CAM has been extended beyond image classification to:

• Visual Question Answering (VQA): Highlighting image regions relevant to textual queries
• Image Captioning: Localizing pixels that influence specific generated words
• Video Action Recognition: Producing spatiotemporal heatmaps across frame sequences These extensions maintain the core gradient-weighting mechanism while adapting the target concept to different output modalities.

Rigorous evaluation of Grad-CAM explanations involves faithfulness metrics such as:

• Deletion/Insertion curves: Measuring prediction change as pixels are removed or added in order of attribution importance
• Cascading randomization: Comparing explanations from progressively randomized model layers to detect insensitivity to model parameters
• Localization error: Quantifying alignment between heatmap maxima and ground-truth bounding boxes These checks are essential for regulatory submissions under FDA Good Machine Learning Practice guidelines.

The foundational visualization technique that Grad-CAM refines. Saliency maps compute the gradient of the class score with respect to the input image, highlighting pixels that most affect the prediction. Unlike Grad-CAM, they operate directly on the input space rather than intermediate feature maps, often producing noisier, less spatially coherent visualizations. They remain useful for fine-grained pixel attribution but lack the class-discriminative localization that Grad-CAM provides.

A variant of standard backpropagation that modifies the gradient flow through ReLU activations. By zeroing out negative gradients during the backward pass, it produces sharper, cleaner visualizations than raw saliency maps. Often combined with Grad-CAM to create Guided Grad-CAM, which element-wise multiplies the Grad-CAM localization map with guided backpropagation output for high-resolution, class-discriminative explanations.

An axiomatic attribution method that satisfies sensitivity and implementation invariance properties that Grad-CAM lacks. It computes feature importance by accumulating gradients along a path from a baseline (e.g., black image) to the input. While model-agnostic and theoretically grounded, it is computationally heavier than Grad-CAM and does not natively produce the coarse localization maps ideal for radiological region-of-interest highlighting.

A decomposition-based explanation method that redistributes the model's output score backward through the network using conservation of relevance. Unlike gradient-based methods like Grad-CAM, LRP does not rely on gradient computation, making it robust to shattered gradients. It produces pixel-level heatmaps with strong theoretical backing but requires layer-specific propagation rules, adding implementation complexity for heterogeneous diagnostic architectures.

A technique that explains neural network decisions in terms of human-friendly concepts rather than individual pixels or feature maps. CAVs test how sensitive a prediction is to a high-level concept (e.g., 'spiculated mass' in mammography) by training a linear classifier on intermediate activations. Complements Grad-CAM by providing semantic, rather than spatial, interpretability for regulatory submissions.

The domain-specific application of attribution methods to identify which biological measurements most influence a diagnostic model's prediction for a specific patient. Extends Grad-CAM's localization philosophy beyond imaging to tabular omics data. Critical for FDA submissions where demonstrating that a model attends to clinically valid biomarkers—not spurious correlations—is essential for regulatory approval.