Background and Objective
Deep video models estimate left-ventricular ejection fraction (EF) from echocardiography with near-expert accuracy, and post-hoc attribution methods (Chefer relevance for transformers, Grad-CAM for CNNs) are increasingly used to verify that models "look at the right place." However, the faithfulness of these explanations, both spatially and temporally, remains unexamined. Since EF is defined by the end-systolic (ES) and end-diastolic (ED) frames, a faithful explanation must localize the left ventricle (space) and the decisive frames (time).
Methods
We fine-tuned two distinct EF regressors on EchoNet-Dynamic: a self-supervised VideoMAE transformer and a Kinetics-pretrained R(2+1)D CNN, and audited each with architecture-matched attribution across three axes: intersection-over-relevance (IoR) against LV masks, deletion AUC, and a temporal localization index on ES/ED frames, each relative to chance with per-case 95% CIs over 50 studies. A tubelet-occlusion probe separates attribution failure from model behavior.
Results
Both models are anatomically faithful—IoR of 2.91x (VideoMAE) and 1.98x (R(2+1)D) above chance—but temporally blind: temporal localization is indistinguishable from chance (0.97–1.00) and no better than random attribution. Occlusion shows the models do not preferentially rely on ES/ED (0.90x chance), thus temporal blindness reflects model behavior, not an attribution artifact.
Conclusions
Spatial faithfulness does not imply temporal faithfulness. Attribution can certify anatomical grounding while masking that a model ignores the clinically decisive frames—this serves as a caution for XAI-based validation of video diagnostic models and a call for temporally-aware training and evaluation.
Blogger's Review: This paper highlights the limitations of deep learning models in medical image analysis, particularly in handling temporal information. While the models perform well spatially, their temporal blindness could lead to clinical decision-making errors, emphasizing the necessity of considering temporal factors in model design and evaluation.