Attention Heatmap

Attention heatmaps assign a saliency score to every spatial location in a whole-slide image (WSI). These scores quantify each region's contribution to the model's final prediction.

• High attention (warm colors): Regions the model considers diagnostically critical, such as tumor nests or dense immune infiltrates.
• Low attention (cool colors): Background stroma, adipose tissue, or glass background the model learns to ignore.
• Resolution: Heatmaps are generated at the patch level (e.g., 256×256 pixels at 20× magnification) and stitched back into a gigapixel map.

Heatmaps serve as a debugging tool for identifying when models rely on spurious correlations rather than genuine pathology.

• Pen mark detection: If attention concentrates on blue pen marks rather than tissue, the model has learned an artifact shortcut.
• Blurry region flagging: Low attention across entire regions may indicate out-of-focus areas the model cannot interpret.
• Tissue fold identification: Heatmaps reveal if the model incorrectly weights folded or torn tissue sections.
• Validation protocol: Pathologists review heatmaps to confirm the model is 'looking' at clinically relevant morphology before trusting its output.

Different attention formulations produce distinct heatmap characteristics, each suited to specific diagnostic tasks.

• Self-Attention (Vision Transformers): Captures long-range dependencies between distant tissue regions, revealing how the model relates stromal patterns to tumor morphology across millimeters of tissue.
• Multi-Head Attention: Multiple parallel attention maps attend to different morphological features simultaneously—one head may focus on nuclear atypia while another tracks immune cell distribution.
• Cross-Attention: Used in Multiple Instance Learning (MIL) to compute compatibility between patch features and a learned query vector representing the diagnostic class.
• Gated Attention: Introduces a learnable gating mechanism that suppresses noisy patches, producing cleaner heatmaps with higher contrast between relevant and irrelevant regions.

Attention heatmaps are increasingly cited in FDA submissions and CE marking documentation as evidence of model interpretability.

• Explainability requirement: Regulatory bodies expect AI/ML-based medical devices to provide interpretable outputs that pathologists can verify.
• Audit trail: Heatmaps create a permanent visual record of which regions influenced a diagnostic decision, supporting retrospective case review.
• Human-AI collaboration: Heatmaps guide pathologists to regions of interest, reducing review time while ensuring the human remains the final decision-maker.
• Limitation disclosure: Submission documents must acknowledge that attention weights indicate correlation, not causation—high attention does not guarantee a region is biologically causative of the diagnosis.

Two dominant computational methods generate heatmaps from different model types, each with distinct properties.

• Grad-CAM (Gradient-weighted Class Activation Mapping): Uses gradients flowing into the final convolutional layer to produce coarse localization maps. Works with any CNN architecture without architectural modification.
• Attention Rollout: Propagates attention weights through all transformer layers by multiplying attention matrices along residual connections, revealing how information flows from input patches to the classification token.
• Key difference: Grad-CAM produces class-discriminative maps (showing what distinguishes one class from another), while attention rollout shows all attended regions regardless of class specificity.
• Hybrid approaches: Combining both methods can validate that class-specific and general attention converge on the same tissue regions.

Beyond model debugging, attention heatmaps enable discovery-driven research by identifying previously unrecognized morphologically predictive regions.

• Novel biomarker identification: If attention consistently highlights a specific stromal pattern not described in current grading guidelines, it may represent a new prognostic feature.
• Genotype-phenotype mapping: Correlating high-attention regions with spatial transcriptomics reveals which gene expression programs drive morphological patterns the model finds predictive.
• Treatment response prediction: Heatmaps from models predicting immunotherapy response can identify whether attention focuses on tumor cells or immune infiltrates, providing mechanistic insight into the prediction.
• Tumor heterogeneity quantification: The spatial distribution and variance of attention scores across a tumor bed quantifies how heterogeneous the predictive morphology is within a single slide.

The neural architecture that makes attention heatmaps possible in modern pathology. ViTs divide a whole-slide image into a sequence of non-overlapping patches, then apply multi-head self-attention to model long-range spatial dependencies between distant tissue regions. The raw attention weights from the final transformer blocks are the direct source data for generating interpretability heatmaps. Unlike convolutional networks that require gradient-based approximations, ViTs provide inherent attention matrices that can be visualized natively.

The dominant weakly-supervised paradigm for whole-slide image classification where only a slide-level label is available. In attention-based MIL, the model learns to assign an attention score to each patch, representing its contribution to the final diagnosis. The attention heatmap is the spatial visualization of these learned weights, revealing which tissue regions the model considers diagnostically relevant. Key variants include:

• ABMIL: Uses a gated attention mechanism for patch aggregation
• CLAM: Cluster-constrained attention for identifying distinct morphological regions
• TransMIL: Combines transformer self-attention with MIL pooling

An alternative interpretability method that uses the gradients of the target class flowing into the final convolutional layer to produce a coarse localization map. Unlike native attention heatmaps from ViTs, Grad-CAM requires backpropagation to compute importance weights for each feature map channel. In pathology, Grad-CAM is commonly applied to CNN-based models and serves as a validation baseline against which attention-based heatmaps are compared. The resulting heatmap highlights discriminative regions but at lower spatial resolution than patch-level attention.

A game-theoretic approach to model interpretability that assigns each input feature an importance value for a particular prediction. In digital pathology, SHAP values can be computed for individual image patches or cellular features to explain why a model classified a slide as malignant. Unlike attention heatmaps that show where the model looked, SHAP explains how much each feature contributed to the output magnitude. This provides complementary information: attention reveals spatial focus, while SHAP quantifies feature impact with formal axiomatic guarantees of consistency and local accuracy.

The high-throughput extraction of quantitative features from the regions identified by attention heatmaps. Once a heatmap highlights diagnostically salient tissue areas, pathomics pipelines extract hundreds of morphological, textural, and spatial descriptors from those regions. These features—including nuclear shape factors, chromatin texture patterns, and stromal architecture metrics—transform qualitative attention into numerical biomarkers suitable for downstream statistical modeling. The heatmap-to-pathomics pipeline bridges visual interpretability with quantitative prognostic model development.