Multimodal Explainability

Computes the partial derivative of a model's output with respect to each input feature across modalities. Saliency maps for images and token importance scores for text are generated simultaneously, revealing which pixels in a chest X-ray and which words in a clinical note most influenced a diagnosis. Integrated Gradients is the preferred variant for satisfying the completeness axiom in multi-modal settings.

Traces the cross-attention weights between modalities in a multimodal transformer. For a radiology report and image pair, this technique visualizes exactly which anatomical regions the text tokens attended to during fusion. Attention rollout and attention flow algorithms aggregate raw attention heads into human-interpretable heatmaps, exposing spurious correlations like a model focusing on a chest drain rather than a pathology.

Extends SHapley Additive exPlanations to quantify each modality's marginal contribution to a prediction. KernelSHAP treats entire modality encoders as features, while DeepSHAP propagates importance through the fusion layers. The output is a unified force plot showing, for example, that a genomic marker contributed 60% and a CT scan contributed 40% to a high-risk classification.

Moves beyond raw features to explain decisions in terms of high-level, human-understandable concepts. TCAV (Testing with Concept Activation Vectors) is adapted to multi-modal data by defining concepts that span modalities, such as 'spiculated mass margin' in imaging and 'invasive ductal carcinoma' in pathology text. The model's sensitivity to each concept is quantified, providing a semantically meaningful explanation.

Generates minimal perturbations to input data that would flip a model's decision. A counterfactual explanation might state: 'If the tumor size in the MRI were reduced by 2mm and the Ki-67 protein expression were below 14%, the diagnosis would change from malignant to benign.' This technique is crucial for clinical actionability, directly informing treatment targets.

Systematically removes or masks one modality at inference time to measure the performance delta. A significant drop in AUC when genomic data is removed indicates high reliance. This technique identifies modality redundancy and complementarity, answering the critical question: 'Is the expensive genomic sequencing strictly necessary for this diagnosis, or does imaging alone suffice?'