Optimal Transport Explanations for Clinical AI: Localization Challenges

Generative models hold promise for creating explainable AI in clinical settings by modeling distributions of healthy and diseased patients and deriving explanations from the geometric differences between them. This paper develops such a system using an optimal-transport rectified flow, trained to map between these two clinical distributions, and then rigorously tests its ability to localize disease. For tabular tumor biomarkers, specifically the Breast Cancer Wisconsin dataset, the system successfully generates per-patient counterfactuals, an unsupervised malignancy score with high AUROC, and label-free attributions that correlate with supervised classifiers. This demonstrates a compact and transparent interpretability engine, though it doesn't surpass logistic regression in predictive power. However, when applied to chest X-rays, a critical "synthetic-to-real gap" emerges. While the transport heatmap provides a population-level signal, it fails to localize disease at the individual patient level. Even a reconstruction-based variant that localizes synthetic lesions collapses to chance on real RSNA radiologist boxes, whereas supervised Grad-CAM maintains its performance. The central finding is that visually compelling label-free heatmaps on synthetic data are not reliable evidence of real disease localization, highlighting a crucial limitation for explainable AI in medical imaging.