Lesion Attribution

The primary function of lesion attribution is to verify anatomical correspondence between the model's attention and the clinical region of interest. This involves:

• Comparing generated saliency maps against radiologist-annotated ground truth segmentations
• Measuring overlap using metrics like Intersection over Union (IoU) and Dice similarity coefficient
• Detecting shortcut learning where models exploit spurious correlations such as scanner type, hospital site markers, or patient positioning rather than pathology

A model that classifies pneumothorax based on chest tube presence rather than the collapsed lung itself exhibits a lesion attribution failure.

A robust lesion attribution system must demonstrate immunity to confounding signals that correlate with disease status but are not themselves pathological. Common confounders include:

• Radiopaque markers and surgical implants that indicate prior intervention
• Scanner-specific noise patterns that leak institution-level information
• Patient demographic features inadvertently encoded in image texture
• Burn-in annotations or laterality markers placed by technicians

Attribution methods like Integrated Gradients with anatomical baseline images help isolate true pathological signal by subtracting out the expected anatomical background before computing feature importance.

Clinical reliability demands convergent validity across multiple attribution algorithms. No single saliency method is universally faithful, so lesion attribution workflows typically employ:

• Grad-CAM for coarse localization of convolutional feature activation
• Integrated Gradients for pixel-level attribution satisfying the completeness axiom
• SHAP for game-theoretic feature importance with strong theoretical guarantees
• Occlusion sensitivity as a model-agnostic perturbation baseline

When all methods independently highlight the same lesion boundary, attribution consensus is achieved. Divergent explanations signal potential model fragility or out-of-distribution inputs requiring clinician review.

In volumetric imaging (CT, MRI), lesion attribution must demonstrate 3D spatial coherence across adjacent slices. A valid attribution exhibits:

• Smooth saliency transitions between consecutive axial, coronal, or sagittal slices
• Attribution that tracks the volumetric extent of the lesion rather than appearing on isolated slices
• Temporal stability in longitudinal studies—the same lesion should be attributed consistently across follow-up scans unless genuine progression has occurred

Jittery or slice-isolated saliency often indicates the model is responding to reconstruction artifacts or slice-level noise rather than true anatomical pathology.

Lesion attribution is not purely a computational metric—it requires expert human judgment for final validation. The standard workflow includes:

• Blinded review where radiologists rate attribution map quality on a Likert scale
• Assessment of whether highlighted regions correspond to clinically actionable findings
• Evaluation of false positive attributions that highlight benign structures as suspicious
• Measurement of trust calibration—whether clinician confidence in the model appropriately tracks attribution quality

This human validation loop is essential for regulatory submissions under FDA SaMD pathways and EU MDR requirements, where explainability must be demonstrated to satisfy GMLP (Good Machine Learning Practice) standards.

Objective evaluation of lesion attribution quality uses perturbation-based and axiom-based metrics implemented in toolkits like Quantus and Captum:

• Faithfulness Correlation: Pearson correlation between feature importance and the actual drop in model confidence when those features are removed
• Monotonicity: Whether adding features in order of attributed importance produces monotonically increasing prediction scores
• Completeness: Whether the sum of all feature attributions equals the difference between the model output and a baseline prediction
• Sensitivity-n: The degree to which explanations change under minimal input perturbations

A clinically deployable system should achieve faithfulness scores above 0.80 across these metrics before integration into diagnostic workflows.