Layer-wise Relevance Propagation (LRP)

LRP operates on a strict relevance conservation axiom. The total relevance score at the output layer (the prediction value) must equal the sum of relevance scores across all input features. This ensures no relevance is created or destroyed during backpropagation.

• Mathematical Guarantee: ∑_i R_i^(l) = ∑_j R_j^(l+1) = f(x)
• Contrast with Gradients: Standard gradients can vanish or explode; LRP maintains a stable signal.
• Significance: This property makes LRP explanations auditable, as the sum of pixel relevances directly maps to the model's confidence score.

LRP is theoretically grounded in Deep Taylor Decomposition (DTD). It approximates the model's decision function locally at each neuron using a first-order Taylor expansion, then redistributes the function value (relevance) backward.

• Mechanism: Identifies root points in the input space where the neuron's activation is minimal.
• Propagation Rules: Different rules (e.g., LRP-αβ, LRP-ε) are derived based on the choice of root point.
• Advantage: Provides a principled mathematical framework rather than relying on heuristic perturbations.

LRP is not a single algorithm but a family of propagation rules tailored to specific layer types or network architectures. Selecting the appropriate rule is critical for explanation fidelity.

• LRP-0: Basic rule, redistributes relevance proportional to positive weights.
• LRP-ε: Adds a small stabilization term to the denominator to dampen noise.
• LRP-αβ: Separates positive and negative contributions, allowing control over the influence of inhibitory evidence.
• LRP-z⁺: Standard choice for deep networks, ignoring negative activations in the redistribution.

Unlike concept-based or region-based methods, LRP produces fine-grained heatmaps that attribute relevance to individual input features (e.g., single pixels in an image or specific nucleotides in a sequence).

• Clinical Utility: In digital pathology, LRP can highlight the exact cellular morphology driving a tumor classification.
• Sparsity: Explanations tend to be sparse, focusing only on the true causal drivers.
• Comparison: Offers higher resolution than Grad-CAM, which produces coarse localization maps limited by the final convolutional layer size.

LRP explicitly models positive relevance (evidence for a class) and negative relevance (evidence against a class). This dual-channel output is crucial for understanding contradictory signals in diagnostic models.

• Red Pixels: Features that increase the predicted probability.
• Blue Pixels: Features that decrease the predicted probability (contradictory evidence).
• Regulatory Relevance: FDA reviewers can assess if a model is using clinically valid biomarkers or spurious correlations by analyzing the sign of the relevance.

The propagation process is strictly layer-wise, meaning relevance is redistributed from layer l+1 to layer l using only the activations and weights of those two adjacent layers. This locality constraint ensures computational tractability.

• Scalability: Can be applied to very deep architectures without requiring a full backward pass through the computational graph.
• Modularity: Allows practitioners to inspect intermediate relevance maps to debug internal representations at any depth.
• Contrast: Global methods like SHAP often require sampling-based approximations that ignore the network's layered topology.

LRP generates heatmaps that highlight the exact pixels in an MRI or CT scan influencing a model's diagnosis. This allows radiologists to verify that a convolutional neural network is focusing on pathological structures like tumors or lesions, not on imaging artifacts or background tissue. The technique is crucial for building trust in computer-aided detection (CADe) systems.

In digital pathology, LRP decomposes predictions on gigapixel whole-slide images to identify which cellular and tissue morphologies drive a cancer classification. By propagating relevance from a 'malignant' output back through a deep network, pathologists can see if the model correctly identifies mitotic figures and nuclear atypia rather than pen marks or staining inconsistencies.

For time-series diagnostic models, LRP reveals which segments of a cardiac waveform are most relevant for detecting conditions like atrial fibrillation or myocardial infarction. The technique propagates relevance backward through recurrent or temporal convolutional layers, highlighting the specific P-wave, QRS complex, or ST-segment deviations that triggered the classification.

When deep learning models classify patients based on gene expression profiles, LRP assigns relevance scores to individual genes. This identifies a sparse set of biomarker candidates that mechanistically drive the prediction, distinguishing true disease-associated genes from correlated but non-causal transcripts. The approach supports target identification in precision oncology.

LRP can be extended to models that integrate imaging, genomic, and electronic health record (EHR) data. By tracing relevance through separate input branches, clinicians can understand the relative contribution of each modality to a final prognosis. This is essential for validating that a model's risk stratification is based on a holistic, clinically sound assessment rather than a single spurious variable.

The FDA and other regulatory bodies increasingly expect explainability for Software as a Medical Device (SaMD). LRP provides the fine-grained, quantitative evidence needed for Predetermined Change Control Plans (PCCPs). By demonstrating that a model consistently relies on clinically valid features across diverse patient subgroups, developers can build a stronger case for safety and effectiveness during the 510(k) or De Novo clearance process.