Detection of Aortic Valve from Echocardiography in Real-Time Using Convolutional Neural Network

This study proposes the development of Convolutional Neural Network (CNN) for an automatic detection system using various deep learning methods for echocardiography. We tested the performance of the CNN by measuring accuracy and the effect on video frame rate to best represent the application of CNN in an actual patient echocardiography examination. The study focuses on the system detecting the aortic valve of the heart as it is clinically significant. Single Shot Multibox Detector (SSD) and Faster Regional based Convolutional Neural Network (R-CNN) architectures with various feature extractors were trained on echocardiography images from 23 patients. Afterward, the detection models were tested on 5 echocardiography videos. The results showed that the Faster R-CNN Inception v2 attained the best accuracy (0.949) and F1 score (0.950). The second-best performer was SSD Inception v2 with 0.865 accuracy and 0.844 F1 score. In terms of prediction speed, SSD architectures were relatively faster and achieved mean frame rate of 34.22 frames-per-second (fps) and 27.66fps for MobileNet and Inception v2 feature extractor respectively. However, the frame rate performance loss for SSD Inception v2 was 49.71% compared to the original 55fps echocardiography video. The findings in this study facilitate a foundation in utilizing convolutional neural network to the echocardiography field.