Anatomically Faithful but Temporally Blind: Auditing Attribution for Left-Ventricular Ejection-Fraction Estimation from Echocardiography

This study investigates whether deep learning models, which estimate left-ventricular ejection fraction (LVEF) from echocardiogram videos with high accuracy, actually rely on the clinically relevant information to make their predictions. Specifically, the authors examine if attribution methods—tools used to explain model decisions—are faithful to both the spatial anatomy (the left ventricle) and the temporal phases (end-systolic and end-diastolic frames) that define LVEF. The researchers audited two distinct model architectures: a self-supervised VideoMAE transformer and an R(2+1)D convolutional neural network, using architecture-matched attribution methods (Chefer relevance and Grad-CAM, respectively).

The audit reveals a significant dissociation between spatial and temporal faithfulness. Both models demonstrate strong spatial grounding, with attribution maps concentrating on the left ventricle at rates significantly higher than chance. However, both models are temporally blind; their attribution maps show no preference for the end-systolic (ES) or end-diastolic (ED) frames, performing no better than random chance. To determine if this was a failure of the explanation methods or the models themselves, the authors performed a tubelet-occlusion probe. This revealed that the models do not actually rely on the ES/ED frames to calculate EF, confirming that the temporal blindness is an accurate reflection of the models' internal decision-making behavior.

These results provide a critical warning for the clinical deployment of AI in medical imaging. While attribution maps may look convincing because they highlight the correct anatomy, they can mask the fact that the model is ignoring the most clinically decisive temporal information. This suggests that current XAI-based validation is insufficient for ensuring that models are making decisions for clinically valid reasons. The authors argue that future efforts must shift toward developing temporally-aware training and evaluation protocols to ensure models are truly learning the physiological basis of cardiac function.