Deep learning-based approach for detecting signs of atrial septal defect on chest radiographs: a proof of concept study

Abstract

AbstractMany patients with atrial septal defects (ASD) are asymptomatic and undiagnosed during the first few decades of life, but have overt heart failure, arrhythmias, cerebral infarction, and increased mortality in adults with advancing age. To provide a non-invasive, easy-to-use, and effective method for detecting ASD, we aimed to develop and validate a deep learning-based algorithm to diagnose ASD on chest radiographs. The ASD dataset was created from 173 chest radiographs of 74 patients with ASD and 170 chest radiographs of 100 patients without ASD. Convolutional neural network models (VGG16, ResNet50, DenseNet121, and Xception) for diagnosing ASD were pretrained using two different datasets, the large-scale real-world ImageNet dataset and the ChestX-ray14 dataset released by National Institutes of Health, followed by a round of training using the training set of the ASD dataset. Model performance was evaluated by five-fold stratified cross-validation. The best performance in ImageNet pretraining was achieved by ResNet50 model, and the cross-validation area under the curve (AUC) was 0.95, with sensitivity of 0.86, specificity of 0.87, and overall accuracy of 0.87. The best performance in ChestX-ray pretraining was achieved by Xception, and the cross-validation AUC was 0.93, with sensitivity of 0.85, specificity of 0.85, and overall accuracy of 0.85. The diagnostic performances of these models were comparable to those of cardiologists. Gradient-weighted Class Activation Mapping showed that the ImageNet-pretrained model focused on bilateral hilar regions, while the ChestX-ray14-pretrained model focused on areas around cardiac silhouette and lower lung fields. Our deep learning-based algorithms made a diagnosis of ASD on the input chest radiographs with high accuracy, and had potential to help clinicians make accurate diagnosis of ASD from routine chest radiography, leading to improvement of prognosis and quality of life in patients with ASD.