A Joint Mask-Augmented Adaptive Scale Alignment and Periodicity-Aware Method for Cardiac Function Assessment in Healthcare Consumer Electronics

Artificial intelligence (AI)-based medical consumer electronics have attracted increasing attention in the fields of precision medicine and intelligent diagnosis. Accurately predicting echocardiographic parameters, such as the size of the left ventricular (LV) area and the ejection fraction (EF), is crucial for the precise diagnosis and treatment of cardiovascular diseases. However, limited data annotations are typically available for medical data, leading to substantial disparities among the segmentation accuracies obtained by deep learning models across different frames, and varying degrees of overfitting are observed in EF prediction tasks. In this work, a joint mask-augmented adaptive scale alignment (ASA) and periodicity-aware EF prediction method is proposed for cardiac function assessment. This approach enhances the accuracy and reliability of LV segmentation and EF prediction. First, an enhanced strategy is proposed to optimize LV segmentation to address the considerable segmentation accuracy variations across different frames. Second, variants of the ASA method are presented to perform regional alignment in EF prediction tasks, and an existing dataset is extended to address model overfitting. Third, a periodicity-aware method is proposed for EF and keyframe prediction, correcting the unfocused attention of the model for different periods and frames within the same video. Finally, experiments conducted on the EchoNet-Dynamic dataset show that our method achieves segmentation accuracy of 91.51% at end-diastole (ED) and 93.55% at end-systole (ES), surpassing previous methods by 1.12% and 0.59%, respectively. The mean absolute error (MAE) in the EF prediction task is 3.92, and the coefficient of determination (R2) is 0.822, demonstrating that our approach outperforms the current state-of-the-art methods.