Deep learning algorithm for detection of aortic dissection on non-contrast-enhanced CT

Objectives To develop a deep learning–based algorithm to detect aortic dissection (AD) and evaluate the diagnostic ability of the algorithm compared with those of radiologists. Methods Included in the study were 170 patients (85 with AD and 85 without AD). An AD detection algorithm was developed using a convolutional neural network with Xception architecture. Of the patient data, 80% were used for training and validation and 20% were used for testing. Fivefold cross-validation was performed to evaluate the method. An average of 6688 non-contrast-enhanced CT images (slice thickness, 5 mm) were used for training. A radiologist reviewed both contrast-enhanced and non-contrast-enhanced images and identified the slices of AD. The identified slices were used as ground truth. Receiver operating characteristic curve and area under the curve (AUC) analysis was performed. Five radiologists independently evaluated the images. The accuracy, sensitivity, and specificity of the algorithm and those of the radiologists were compared. Results The AUC of the developed algorithm was 0.940, and a cutoff value of 0.400 provided accuracy of 90.0%, sensitivity of 91.8%, and specificity of 88.2%. For the radiologists, median (range) accuracy, sensitivity, and specificity were 88.8 (83.5–94.1)%, 90.6 (83.5–94.1)%, and 94.1 (72.9–97.6)%, respectively. There was no significant difference in performance in terms of accuracy, sensitivity, or specificity between the algorithm and the average performance of the radiologists ( p > 0.05). Conclusions The developed algorithm showed comparable diagnostic performance to radiologists for detecting AD, which suggests the potential of the proposed method to support clinical practice by reducing missed ADs. Key Points • A deep learning–based algorithm for detecting aortic dissection was developed using the non-contrast-enhanced CT images of 170 patients. • The algorithm had an AUC of 0.940 for detecting aortic dissection. • The accuracy, sensitivity, and specificity of the algorithm were comparable to those of radiologists.