Intelligent diagnosis of atrial septal defect in children using echocardiography with deep learning

Congenital heart defects (CHDs) are the most common birth defects, occurring in approximately 0.9% of live births. Sadly, they are also the leading cause of death among children 0–5 years old. Approximately 0.15 million newborns are diagnosed with congenital heart disease in China every year, and without treatment, approximately one-third of newborns with congenital heart disease die because of serious complications within the first year of their life. Atrial septal defect (ASD) is one of the most common CHDs in children, accounting for approximately 6%–10% of CHDs, with an approximate population incidence of 1–3 per 1000. The atrial septum is a partition that separates the left atrium and the right atrium during embryonic development. The abnormal formation of this gap may result in a gap after birth. If the opening between the atria fails to close naturally during growth and development, an ASD occurs. With regard to its occurrence, ASDs can be divided into two categories: primary ASDs and secundum ASDs. In this study, secundum ASDs were examined, which can be divided into four types according to the defect location: central defect, superior cavity defect, inferior cavity defect, and mixed defect. Most children with isolated ASDs are free of symptoms, but the rates of exercise intolerance, arrhythmia, right ventricular failure, and pulmonary hypertension increase with age. Therefore, accurate diagnosis of ASD in early childhood is of considerable clinical importance.

Transthoracic echocardiography (TTE) can provide sufficient structural information about the heart and can be used to assess hemodynamics and cardiac function, serving as the effective method for ASD examination. Additionally, ASD can be accurately diagnosed by analyzing the left and right atrial shunt signals obtained through color Doppler flow imaging in echocardiography. The advantages of TTE include simplicity, non-invasiveness, cost-effectiveness, repeatability, and accuracy. However, accurate diagnosis of ASD via TTE relies heavily on the subjective and time-consuming judgments of echocardiographers. In China, owing to the shortage of experienced echocardiographers at the grassroots level, the accurate diagnosis of ASD depends on the individual technical skills of the operator. In addition, low-quality and noisy echocardiograms may lead to missed ASD diagnoses and misdiagnoses, as echocardiograms are affected by the inherent imaging quality of the machine.

With the advancements in artificial-intelligence (AI) technology in recent years, deep-learning methods based on convolutional neural networks (CNNs) are increasingly being applied to various medical image analysis tasks, such as image recognition, organ segmentation, and disease diagnosis. Kusunose proposed four key steps for AI application in echocardiography: image quality preservation acquisition, view classification, measurement and quantification, and abnormality detection. Wu et al. proposed a deep learning-based neural-network method to automatically and efficiently identify the standard echocardiographic view, and the F1-scores for a majority of the views were >0.90. Zhang et al. proposed, trained, and evaluated CNN models for multiple tasks, including automated identification of 23 viewpoints and segmentation of cardiac chambers across five common views. The results indicated that the CNNs accurately identified views and segmentation of the heart. Nevertheless, previous research on standard view recognition and ventricular segmentation cannot satisfy the requirements of accurate diagnosis of CHDs. There is little research on the use of deep-learning systems to detect ASDs. Matsuoka et al. introduced deep learning-based algorithms to detect signs of ASD in chest radiographs. Chen et al. proposed a fully connected network with a multi-scale dilation convolution module for evaluating ASD through magnetic resonance imaging (MRI) segmentation. Li et al. proposed an AI-assisted diagnostic method for identifying ASDs in echocardiograms. However, the paper did not provide a comprehensive evaluation of the model. Hong et al. proposed an automatic detection method for identifying secundum ASDs in children using CNN models, but segmented images were used to train the model, which is laborious and time-consuming.

In clinical practice, ultrasound doctors select a frame of a color Doppler image as the key frame for diagnosing ASD by scanning the cross-sectional view. This is because the color Doppler image can clearly show the blood-flow signal of the interatrial septum level, which is consistent with the color Doppler images used in this study. Therefore, this paper proposes an end-to-end deep-learning method for automatic recognition of ASD based on color Doppler images, which addresses the problems of time-consuming and labor-intensive image segmentation for training the model.