Abstract:
Congenital heart disease (CHD) is a prevalent condition affecting infants and children worldwide, and it is the most common type of birth defect in China. But the shortage of pediatric doctors and the uneven distribution of medical resources in China have led to the lack of an effective screening system for CHD. As a result, many CHD patients cannot receive timely diagnosis and treatment, which significantly affects the quality of the population and increases social healthcare costs. In response to this situation, we rely on new artificial intelligence technology to tailored to our national conditions. Specifically, we propose that CHD screening should be carried out by primary care physicians in community hospitals who are responsible for routine physical examinations using AI-assisted CHD ultrasound model, thereby helping to prevent and control CHD in China. To alleviate this problem in our CHD diagnosis task, the Depth-wise Separable Convolution (DSC) is adopted to largely reduce the network parameters. The decision is based on five views, which can incorporate more complementary information than a single view-based diagnosis. However, this also introduces the challenge of information fusion. A five-channel CNN, as shown in this figure, is developed to take the concatenated matrix as input. Since keyframe selecting always costly, the automatic analysis of multi-view videos in an end-to-end fashion is necessary when large scale samples are collected. Therefore, we propose to investigate the possible deep aggregation frameworks for our video-based multi-view 2D echocardiogram analysis. The basic idea is to assign a larger weight for the more important frame in the diagnosis procedure. In addition to using deep learning models to assist with early screening and diagnosis of CHD, we are also committed to using deep learning models to assist in developing clinical management plans for CHD. We developed a novel deep learning-based AI model (deep keypoint stadiometry, DKS) to localize the endpoints of defects, i.e., keypoint of CHD, in multiview Echo accurately and therefore differentiate the transcatheter or surgical closure following the explicit expert knowledge-based clinical rules. To sum up, our researches indicate that the short-term automated machine learning process can partially replace and promote the long-term professional training of primary doctors, improving the primary diagnosis rate of CHD in China, and laying the foundation for early diagnosis and timely treatment of children with CHD.
