Domain-specific Saliency

Saliency maps are constrained using pre-registered anatomical atlases that define organ boundaries, tissue types, and expected spatial relationships. Atlas-based regularization penalizes high saliency values that fall outside anatomically plausible regions, ensuring that a lung nodule detector's explanation highlights pulmonary parenchyma rather than mediastinal structures. Common atlas resources include the MNI152 brain template, Visible Human Project datasets, and organ-specific segmentation priors from TotalSegmentator.

Domain knowledge about expected lesion morphology is encoded as differentiable constraints during saliency generation:

• Compactness constraints penalize scattered, fragmented saliency blobs that don't correspond to cohesive pathological structures
• Boundary smoothness priors enforce that saliency edges align with tissue interfaces visible in the underlying image
• Size priors suppress saliency regions that are implausibly small or large for the target pathology These priors prevent explanations that highlight noise speckle or irrelevant anatomical edges.

Saliency computation incorporates complementary information from non-imaging clinical data to guide attribution:

• Radiology reports provide textual descriptions of finding locations, used to weakly supervise saliency map generation
• Segmentation masks from expert annotators serve as ground truth for training explanation consistency
• Clinical metadata such as patient age, sex, and lab values condition the prior distribution over plausible anatomical regions This fusion ensures explanations respect the full clinical context rather than operating on images in isolation.

Saliency maps are evaluated against known physiological mechanisms of disease presentation:

• Vascular territory constraints in neuroimaging ensure stroke explanations respect arterial supply regions (MCA, ACA, PCA territories)
• Lymphatic drainage patterns guide saliency in oncology imaging to expected nodal involvement regions
• Biomechanical stress distributions inform orthopedic saliency by weighting regions according to load-bearing patterns Violations of these constraints indicate potential model reliance on spurious correlations or dataset artifacts.

Domain-specific saliency is not purely post-hoc; it can be integrated directly into model training through explanation regularization loss terms:

• Anatomical alignment loss minimizes the distance between generated saliency maps and expert segmentations
• Sparsity penalties encourage focused explanations that concentrate on a single contiguous region
• Consistency losses enforce that saliency maps for augmented versions of the same image remain stable This approach produces models whose internal reasoning is inherently aligned with clinical expectations.

Domain-constrained saliency directly supports regulatory explainability requirements under frameworks like FDA's SaMD guidance and EU MDR:

• Provides auditable evidence that model decisions are based on medically relevant image regions
• Enables automated flagging of explanations that violate anatomical constraints, triggering human review
• Supports clinician-in-the-loop workflows where saliency maps are overlaid on anatomical references for rapid verification
• Contributes to SaMD audit trails by logging whether each prediction's explanation satisfied domain validity checks

Constrains saliency map generation using a pre-registered anatomical atlas as a spatial prior. The model's attention is penalized if it falls outside physiologically plausible regions for a given pathology.

• Mechanism: A regularization term added during training or post-hoc that compares the saliency map to a binary organ mask.
• Example: For a lung nodule classifier, the saliency is constrained to the lung parenchyma, ignoring ribs or mediastinum.
• Benefit: Eliminates the interpretability illusion where a model appears to focus on the lesion but is actually using a confounding chest tube.

A validation protocol that quantifies the overlap between a model's saliency map and the radiologist-annotated ground-truth lesion boundary.

• Metric: Dice similarity coefficient between the saliency hotspot and the segmentation mask.
• Clinical Significance: Directly proves the model is 'looking at' the pathology, a core requirement for FDA SaMD clearance.
• Workflow: Integrated into the clinician-in-the-loop review, flagging cases where the Dice score drops below a threshold for manual inspection.

A training-time technique that adds a penalty to the loss function to enforce topological and morphological constraints on the resulting saliency maps.

• Constraints: Encourages saliency to be compact, contiguous, and match the expected shape of the target structure (e.g., spherical for a nodule, tubular for a vessel).
• Implementation: Uses a differentiable loss term based on the saliency map's center of mass and moment of inertia.
• Outcome: Prevents fragmented or scattered explanations that erode trust calibration with radiologists.

Aggregates saliency maps generated under the guidance of multiple probabilistic anatomical atlases to account for inter-patient anatomical variability.

• Process: The input image is non-rigidly registered to a library of atlases. A saliency map is generated for each, and the final map is a voxel-wise average.
• Robustness: Reduces the risk of a single atlas mismatch causing a false-negative explanation.
• Application: Critical in neuroimaging where cortical folding patterns vary significantly between individuals.

Extends saliency beyond spatial regions to functional physiological pathways, such as the cardiac conduction system or a specific white matter tract.

• Method: Uses a graph neural network where nodes are anatomical regions and edges are known functional connections. Attribution is performed on the graph edges.
• Example: For an ECG-based arrhythmia classifier, the explanation highlights the sinoatrial node to atrioventricular node pathway rather than individual ECG lead pixels.
• Regulatory Impact: Aligns explanations with the mechanistic reasoning clinicians use, satisfying regulatory explainability requirements.

Combines uncertainty attribution with anatomical priors to show not just where the model is looking, but how confident it is in that focus, within a clinical context.

• Visualization: Overlays a heatmap where color intensity represents saliency and opacity represents epistemic uncertainty.
• Clinical Decision Support: A bright but transparent region signals 'the model is focusing here but is unsure,' prompting the clinician to investigate further.
• Goal: Directly supports trust calibration by preventing over-reliance on high-saliency but high-uncertainty regions.