A Deep-Learning-based Cardiac Sound Segmentation Method for Smart Auscultation Applications

Automatic phonocardiogram (PCG) segmentation is important for smart cardiac auscultation. One cardiac sound cycle includes the first heart sound (S1), systole, the second heart sound (S2), and diastole. Due to individual variation or pathological changes, PCG may show fixed or unfixed length of interval, retaining or vanishing of S1 or S2, or various abnormal components. E. g., systolic murmurs in aortic stenosis (AS) may also lead to disappearance of S2, hindering locating S1 and S2 in traditional cardiac sound segmentation methods. In this paper, based on our established R-CNN-style deep-learning model, a novel PCG cycle segmentation method is proposed. Firstly, multi-scale temporal region proposals are extracted from Shannon-entropy envelope. Secondly, a convolutional neural network (CNN) for classification and regression of temporal region proposals is established, where the backbone network is DenseNet with attention mechanism. Finally, non-maximum suppression (NMS) algorithm is employed to ensure the most accurate result from the PCG segments of temporal region proposals. The proposed method has been proved robust on multiple databases, including 95.92% accuracy on our own database, 95.52% on PhysioNet 2016, and 92.32% on PASCAL. Specifically, the proposed method can achieve 90.41% accuracy on AS PCG recordings.