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

2026-07-15

AI Summary

Current deep video models for estimating left ventricular ejection fraction (EF) from echocardiography lack systematic validation of the spatiotemporal fidelity of their post-hoc attribution methods. This study conducts a multidimensional audit of attributions from VideoMAE and R(2+1)D models using anatomical mask Intersection-over-Union, deletion AUC, and a temporal localization index, complemented by a tubelet masking probe to distinguish attribution failures from genuine model behavior. The work reveals, for the first time, that while spatial attributions significantly outperform random baselines (IoR 1.98–2.91×), temporal localization performance is statistically indistinguishable from random (0.97–1.00×), and models do not preferentially rely on clinically critical end-systolic or end-diastolic frames. These findings suggest that strong spatial fidelity may mask underlying temporal deficiencies, challenging prevailing XAI validation paradigms and underscoring the necessity of independent assessment of temporal attribution reliability.

Abstract

Background and Objective: Deep video models estimate left-ventricular ejection fraction (EF) from echocardiography with near-expert accuracy, and post-hoc attribution (Chefer relevance for transformers, Grad-CAM for CNNs) is increasingly used to certify that models "look at the right place." Yet whether these explanations are faithful both spatially and temporally is unaudited. Because 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-tune two distinct EF regressors on EchoNet-Dynamic -- a self-supervised VideoMAE transformer and a Kinetics-pretrained R(2+1)D CNN -- and audit each with architecture-matched attribution along 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 2.91x (VideoMAE) and 1.98x (R(2+1)D) above chance -- yet 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), so 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 -- a caution for XAI-based validation of video diagnostic models and a call for temporally-aware training and evaluation.

Problem

• left-ventricular ejection fraction
• echocardiography
• temporal localization
• attribution faithfulness
• video diagnostic models

Innovation

• temporal faithfulness
• attribution auditing
• left-ventricular ejection fraction
• video diagnostic models
• tubelet occlusion