Automated Detection and Differentiation of Stanford Type A and Type B Aortic Dissections in CTA Scans Using Deep Learning

Figure 1. Example of aortic CTA in our study. (A) showed aortic dissection, and (AD) involved the ascending and descending aorta, which was the typical imaging presentation of Stanford type A AD. (B) displayed an intimal flap in the descending aorta, suggesting a Stanford type B AD. (C) was a normal case.

Figure 2. Diagram shows deep learning algorithms. An objection detection model would identify the ascending aorta, aortic arch, and descending aorta from CTA scan. These cropped patches were then sent to the classification model to determine the presence of aortic dissection. After classification, the cropped patches were arranged in a sequence from ascending to descending aorta to determine their Stanford type. The “P.” and “N.” labels under each cropped patch indicate whether the patch is positive or negative for the presence of dissection.

Figure 3. Flowchart shows participants selection. Among the 2243 patients with aortic CTA scans from 2018 to 2020, 1718 patients were included if the following criteria was met: (1) clinical presentation of atypical chest pain, (2) an aortic CTA for evaluating aortic dissection was performed and (3) age over 20 years. A total of 903 patients were excluded because of (1) missing data or medical record, (2) suboptimal CTA scans, (3) presence of intra-aortic device or surgical change, (4) thrombus in the aorta and (5) other co-existing findings. A total of 814 patients were enrolled. A figure of 498 patients were randomly assigned for model development, and 316 patients constitute the independent test set.

Figure 4. The 3 × 3 confusion matrix. The case numbers of the whole system’s prediction and imaging diagnoses, obtained from the independent test set, were demonstrated in the 3 × 3 confusion matrix.