Hardware-aware Real-time Myocardial Segmentation Quality Control in Contrast Echocardiography
Event Type
Special Session (Research Track)
Virtual Programs
Hosted in Virtual Platform
Machine Learning/AI
DescriptionAutomatic myocardial segmentation of contrast echocardiography plays a critical role in the quantification of myocardial perfusion parameters. Conventionally, during the data acquisition process (ultrasound scan), the image quality is controlled by the operating technician based on human vision only. In this talk, we will for the first time demonstrate that interestingly, high quality images from human vision perspective do not necessarily lead to good segmentation results by machine learning algorithms. It is therefore imperative to predict the segmentation quality on-the-fly during the ultrasound scan. However, it is infeasible to deploy state-of-the-art DNN-based models because they are too large (even after compression) to fit in the limited hardware resource on the ultrasound machine and too slow to satisfy the strict real-time latency constraints. In this talk, we will introduce a novel hardware-aware compact neural network scheme for quality control of data acquisition in contrast echocardiography. Experiments on over 100 of our patients show the proposed quality control scheme can achieve more than 50% reduction in segmentation failures that require human correction, compared with the conventional human-vision only scheme.