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 highlighting the discriminative regions used by the CNN for prediction. This mechanism makes it broadly applicable to any CNN architecture without architectural changes.

Unlike Class Activation Mapping (CAM), which requires a specific architecture ending with a Global Average Pooling layer followed by a single fully-connected layer, Grad-CAM is model-agnostic. It can be applied to any convolutional neural network, including:

• ResNet variants
• Inception networks
• DenseNet architectures
• Vision Transformers with convolutional components This flexibility makes Grad-CAM the standard for post-hoc visual explanation in medical imaging pipelines where diverse backbone architectures are common.

Standard Grad-CAM produces low-resolution heatmaps due to the spatial downsampling in deep convolutional layers. To achieve pixel-level granularity, Grad-CAM is often fused with guided backpropagation to create Guided Grad-CAM. This hybrid approach multiplies the upsampled Grad-CAM localization map with the guided backpropagation visualization, yielding sharp, high-resolution saliency maps that precisely delineate lesion boundaries in radiological scans.

Grad-CAM can be extended to generate counterfactual explanations by negating the gradient of the target class. This produces a heatmap highlighting the regions that, if removed or altered, would decrease the model's confidence in its prediction. In medical object detection, counterfactual Grad-CAM reveals which anatomical structures the model relies on, helping clinicians verify that the model is attending to pathological regions rather than spurious correlations like scanner artifacts or patient positioning.

Grad-CAM supports class-discriminative localization for any neuron in the network, not just the final classification layer. This enables multi-task visualization where separate heatmaps are generated for:

• Different pathology classes (e.g., nodule vs. mass vs. consolidation)
• Intermediate semantic concepts learned by hidden layers
• Bounding box regression branches in detection architectures This granularity is critical for debugging object detection models in radiology, where a single model must simultaneously localize and classify multiple abnormality types.

The faithfulness of Grad-CAM explanations is evaluated using perturbation-based metrics and deletion/insertion curves:

• Pointing Game: Measures whether the maximum point of the heatmap falls within the ground-truth bounding box annotation
• Deletion Metric: Quantifies the drop in classification confidence as pixels are progressively removed in order of importance
• Insertion Metric: Measures the rise in confidence as salient pixels are incrementally revealed These metrics provide rigorous, reproducible benchmarks for comparing explainability methods in diagnostic AI.

A visualization technique that modifies the standard backpropagation algorithm by zeroing out negative gradients during the ReLU backward pass. This prevents neurons that decrease the activation of a higher layer unit from contributing to the visualization, resulting in sharper, less noisy saliency maps than standard gradient-based methods. When combined with Grad-CAM to form Guided Grad-CAM, it produces high-resolution, class-discriminative visualizations that highlight fine-grained features like lesion boundaries.

A decomposition-based explanation method that redistributes the prediction score backward through the network using conservation of relevance—the total relevance at each layer remains constant. Unlike gradient-based methods, LRP defines propagation rules tailored to specific layer types:

• LRP-ε: Adds a small stabilizer to prevent division by zero
• LRP-αβ: Separates positive and negative contributions for asymmetric treatment LRP is particularly valued in medical imaging because its explanations are theoretically grounded and pixel-level attributions sum exactly to the model's output score.

A unified framework based on cooperative game theory that assigns each input feature an importance value for a particular prediction. SHAP values satisfy three desirable properties:

• Local accuracy: Explanations sum to the model output
• Missingness: Missing features get zero attribution
• Consistency: If a feature's contribution increases, its SHAP value doesn't decrease For image models, Deep SHAP combines SHAP with DeepLIFT to efficiently approximate Shapley values, providing a theoretically rigorous alternative to Grad-CAM for regulatory-facing explanations.

A technique that sharpens gradient-based saliency maps by adding Gaussian noise to the input image multiple times and averaging the resulting sensitivity maps. The core insight: while a standard saliency map may appear visually noisy due to local gradient fluctuations, averaging over noisy samples smooths out these variations and reveals the genuinely important regions. SmoothGrad is model-agnostic and can be applied on top of Grad-CAM to produce cleaner, more interpretable heatmaps for radiologist review.

An axiomatic attribution method that satisfies sensitivity and implementation invariance—two properties that standard gradients lack. It computes feature importance by accumulating gradients along a straight-line path from a baseline image (typically black or blurred) to the actual input. The integral approximates the continuous change in output as features scale from absent to present. For medical imaging, Integrated Gradients provides theoretically sound attributions that are robust to model architecture choices, making them suitable for FDA submission documentation.

A perturbation-based method that systematically slides a gray patch across the input image and measures the change in the model's predicted probability for the target class. Regions that cause a large drop in confidence when occluded are deemed important. While computationally expensive compared to gradient methods, occlusion analysis is model-agnostic and provides a direct, intuitive measure of spatial importance. It serves as an excellent sanity check for Grad-CAM results—if both methods highlight the same lesion area, confidence in the explanation increases.