Deep learning algorithms for detecting acute thoracic aortic dissection on plain chest radiography: a retrospective multicenter study

Aortic dissection is one of the most life-threatening acute aortic syndromes in which blood leaking from the damaged inner layer of the aorta causes dissection between the intimal and adventitial layers. With the recent development of deep learning technology for image recognition, we developed a deep learning algorithm for screening aortic dissection through chest X-ray scans using a convoluted neural network and evaluated the diagnostic ability of the developed algorithm. The chest X-ray images were obtained from three tertiary academic hospitals. After learning using residual neural network 18- and 5-fold cross-validation with chest X-ray images obtained from two hospitals, a test was performed with data from the remaining one hospital. To validate the performance of five models trained through 5-fold cross-validation, accuracy, precision, recall, and F-1 score were calculated. A total of 3,331 images containing 716 positive images and 2615 negative images were collected from 3,331 patients. Overall, 1,972 images consisting of 507 positive images (male, 62.7%; age [SD], 61 [15] years) in hospital A, 1,155 images consisting of 155 positive images (male, 56.1%; age [SD], 63 [13] years), and 204 images consisting of 54 positive images (male, 55.6%; age [SD], 61 [17] years) were analyzed. The diagnostic accuracy of the deep learning model was 90.20% with precision 75.00%, recall 94.44%, and F1-score 83.61%. In conclusion, the interpretation of chest X-ray images using the CNN algorithm that we developed to detect aortic dissection could help doctors screen patients with suspected aortic dissection.