Deep Learning Based Automatic Segmentation of the Thoracic Aorta from Chest Computed Tomography in Healthy Korean Adults.

ABSTRACT

OBJECTIVE: Segmenting the aorta into zones based on anatomical landmarks is a current trend to better understand interventions for aortic dissection or aneurysm. However, comprehensive reference values for aortic zones are lacking. The aim of this study was to establish reference values for aortic size using a fully automated deep learning based segmentation method. METHODS: This retrospective study included 704 healthy adults (mean age 50.6 ± 7.5 years; 407 [57.8%] males) who underwent contrast enhanced chest computed tomography (CT) for health screening. A convolutional neural network (CNN) was trained and applied on 3D CT images for automatic segmentation of the aorta based on the Society for Vascular Surgery/Society of Thoracic Surgeons classification. The CNN generated masks were reviewed and corrected by expert cardiac radiologists. RESULTS: Aortic size was significantly larger in males than in females across all zones (zones 0 - 8, all p < .001). The aortic size in each zone increased with age, by approximately 1 mm per 10 years of age, e.g., 25.4, 26.7, 27.5, 28.8, and 29.8 mm at zone 2 in men in the age ranges of 30 - < 40, 40 - < 50, 50 - < 60, 60 - < 70, and ≥ 70 years, respectively (all p < .001). CONCLUSION: The deep learning algorithm provided reliable values for aortic size in each zone, with automatic masks comparable with manually corrected ones. Aortic size was larger in males and increased with age. These findings have clinical implications for the detection of aortic aneurysms or other aortic diseases.