Advanced Congenital Heart Disease Diagnosis Based on Automatic Generation of Echocardiogram

Rationale: Evidence show that five views (A4C, SSAX, SXLAX, PSSAX, and PSLAX) of echocardiograms can help to improve the diagnosis accuracy of congenital heart disease, especially with the power of artificial intelligence, which become a popular method in the field of computer-aided diagnosis system. However, there are some problems in clinical trials. For example, compared with the A4C view, obtaining other views have very high requirements for the operator. Therefore, it is not easy to get the exact five views perfectly from a patient. In order to solve this problem, we chose to use the UNet structure to build the generator and discriminator of the GAN network, and use the built StarGAN to generate views that are not easy to obtain. Experiments have found that this method is effective. After that, the generated views are used to supplement the difficult-to-obtain views, and then, the improvement of the classification of congenital heart disease performance can be confirmed. Methods: In order to solve the problem that the diagnostic model cannot be used due to missing views, StarGAN was employed to automatically generate the missing views after screening. After professional doctors’ confirmation, the generated views satisfy the clinical accuracy, which could substitute the missing views. In this way, the data meet the conditions of a high-accuracy DSC diagnostic model, and then DSC is used to diagnose congenital heart disease. Results: In the case that the views generated by StarGAN can be used after the confirmation of professional doctors, it shows that StarGAN can help to solve the problem of missing views. The accuracy of the diagnosis of congenital heart disease supplemented by the generated views is 95.15%, the diagnosis result is negative, VSD or ASD is 92.11%. Conclusions: The technology of automatically generating and complementing other views with the A4C view as the core avoids misdiagnosis caused by incomplete or inaccurate views collected by junior doctors. In addition, the high-accuracy DSC classification model requires five different view inputs. The generated view complements the data, and the diagnosis classification network is trained through five views to assist doctors in diagnosis and treatment, which can be applied to primary hospitals. Funding: This work was supported by the National Natural Science Foundation of China (grant number 91846102), the Fundamental Research Funds for the Central Universities (grant number GK2240260006) and Beijing Hospitals Authority Youth Programme, grant number: QML20191208. Declaration of Interest: None to declare. Ethical Approval: The study protocol was approved by the Ethics Committee of Beijing Children’s Hospital (No. 2019-k-342)