FFANet—Full frequency attention net for automatic diastolic function assessment

Abstract

Using echocardiography to assess the systolic and diastolic function is one of the most important tests of left ventricular function in clinical diagnosis. According to the latest guidelines for assessing diastolic function, the recommended measures and diagnostic methods, such as transmitral flow, tissue velocity, tricuspid regurgitation velocity, and maximum left atrial volume, are complex and rely on multi-angle views. This dilemma leads to there is a significant proportion of patients with left ventricular diastolic dysfunction (LVDD) being difficult to diagnose, which makes it a challenge, especially in the assessment of the patient with heart failure with preserved ejection fraction (HFpEF). In this paper, we aim to simplify the clinical complexity of diastolic function assessment by introducing a full frequency attention net (FFANet) capable of assessing diastolic function using unified morphological cardiac features without processing complex clinical parameters or multi-angle echocardiography images. The proposed method first takes U-Net to segment the left ventricle and left atrium to extract the structure features of apical four-chamber view (A4C) images and then uses convex hull algorithm to transform morphological features of the heart chambers during the dynamic cardiac cycle into one-dimensional linear features. Afterwards, a ResNet-based backbone that embedded full frequency domain attention mechanism instead of traditional global average pooling is employed to model important dynamic morphological features thus assessing diastolic function. The experimental results demonstrate the effectiveness of using echocardiography to assess diastolic function through A4C images independently, thereby simplifying the clinical diagnosis process significantly.