Deep Learning Framework for Interpretable Quality Control of Echocardiography Video

Background: Echocardiography (echo) has become an indispensable tool in modern cardiology, offering real-time imaging that helps clinicians evaluate heart function and identify abnormalities. Despite these advantages, the acquisition of high-quality echo is time-consuming, labor-intensive, and highly subjective.

Purpose: The objective of this study is to introduce a comprehensive system for the automated quality control (QC) of echo videos. This system focuses on real-time monitoring of key imaging parameters, reducing the variability associated with manual QC processes.

Methods: Our multitask network analyzes cardiac cycle integrity, anatomical structures (AS), depth, cardiac axis angle (CAA), and gain. The network consists of a shared convolutional neural network (CNN) backbone for spatial feature extraction, along with three additional modules: (1) a bidirectional long short-term memory (Bi-LSTM) phase analysis (PA) module for detecting cardiac cycles and QC targets; (2) an oriented object detection head for AS analysis and depth/CAA quantification; and (3) a classification head for gain analysis. The model was trained and tested on a dataset of 1331 echo videos. Through model inference, a comprehensive score is generated, offering easily interpretable insights.

Results: The model achieved a mean average precision of 0.962 for AS detection, with PA yielding average frame errors of 1.603 1.181 (end-diastolic) and 1.681 1.332 (end-systolic). The gain classification model demonstrated robust performance (Area Under the Curve > 0.98), and the overall processing speed reached 112.4 frames per second. On 203 randomly collected echo videos, the model achieved a kappa coefficient of 0.79 for rating consistency compared to expert evaluations CONCLUSIONS: Given the model's performance on the clinical dataset and its consistency with expert evaluations, our results indicate that the model not only delivers real-time, interpretable quality scores but also demonstrates strong clinical reliability.