Abstract 275: A New, Novel Shape Feature Space Classifier For Acute Type B Aortic Dissection

Abstract

Background: Early detection and treatment of acute type B aortic dissection (TBAD) is associated with improved morbidity and mortality. Artificial intelligence and machine-learning (ML) based approaches for expedited detection of TBAD have made strides in recent years with improving predictive power and accuracy. These deep-learning neural networks are trained with iterative exposure to patient imaging and clinical information of patients with aortic pathology to identify consistent patterns. AI-decision support tools can be improved with more focused feature selection, variables of particular interest. Here, we identify a novel shape variable to describe the highly complex topography of the aorta. Methods: 46 patients with acute, complicated TBAD were identified from a now closed multi-center clinical trial of a conformable thoracic stent graft. An independent cohort of 87 patients without aortic pathology was obtained from a single center, retrospective database. Two-dimensional surface segmentations were generated from pre-operative CT angiography. Variance in total curvature, a size-independent shape measure, and inverse Casorati curvature, a size scalar, were calculated from the segmentations. Additional size measurements, including maximum diameter, were calculated using centerline methods. Results: 44 pre-operative scans of acute TBAD were analyzed. Mean maximum diameter in the acute TBAD cohort was 39.20 ± 6.75 mm compared to 21.44 ± 3.28 mm in the non-pathologic aorta cohort (p = 2.50E-17). Mean fluctuation in total curvature, normalized, was 7.09 ± 3.58 versus 4.18 ± 2.32 in the non-pathologic aorta cohort (p = 4.36E-13). Mean inverse Casorati curvature was 0.043 ± 0.005 in acute TBAD compared to 0.06 ± 0.01 in non-pathologic aortas (p = 4.67E-21). Logistic regression classification accuracy was 91.6 ± 2.4%. Conclusions: Acute TBAD demonstrates a unique shape signature defined by fluctuation in total curvature, in addition to traditional size measures. This easily quantifiable and reproducible dimension of aortic geometry classification has significant potential as a new, anatomically derived feature space to improve AI/ML-based aortic dissection detection.