Abstract
IntroductionMachine learning techniques have shown excellent performance in 3D medical image analysis, but have not been applied to acute uncomplicated type B aortic dissection (auTBAD) utilizing SVS/STS-defined aortic zones. The purpose of this study was to establish a trained, automatic machine learning aortic zone segmentation model to facilitate performance of an aortic zone volumetric comparison between auTBAD patients based on rate of aortic growth.MethodsPatients with auTBAD and serial imaging were identified. For each patient, imaging characteristics from two CT scans were analyzed: (1) the baseline CTA at index admission, and (2) either the most recent surveillance CTA, or the most recent CTA prior to an aortic intervention. Patients were stratified into two comparative groups based on aortic growth: rapid growth (diameter increase ≥5mm/year) and no/slow growth (diameter increase <5mm/year).Deidentified images were imported into an open-source software package for medical image analysis and randomly partitioned into training(80%), validation(10%), and testing(10%) cohorts. Training datasets were manually segmented based on SVS/STS criteria. A custom segmentation framework was used to generate the predicted segmentation output and aortic zone volumes.ResultsOf 59 patients identified for inclusion, rapid growth was observed in 33 (56%) patients and no/slow growth was observed in 26 (44%) patients. There were no differences in baseline demographics, comorbidities, admission mean arterial pressure, number of discharge antihypertensives, or high-risk imaging characteristics between groups (p>0.05 for all). Median duration between baseline and interval CT was 1.07 years (IQR 0.38-2.57). Post-discharge aortic intervention was performed in 13 (22%) of patients at a mean of 1.5±1.2 years, with no difference between groups (p>0.05). In both groups, all zones of the thoracic and abdominal aorta increased in volume over time, with the largest relative increase in Zone 5 with a median 24% increase (IQR 4.4-37). Baseline zone 3 volumes were larger in the no/slow growth (6v3) than the rapid growth group (5v3) (p=0.03). There were no other differences in baseline zone volumes between groups (p>0.05 for all). Dice coefficient, a performance measure of the model output, was 0.73. Performance was best in Zones 4 (0.82), 5(0.88), and 9(0.91).ConclusionsTo our knowledge this is the first description of an automatic deep learning segmentation model incorporating SVS-defined aortic zones. The open-source, trained model demonstrates high concordance to the manually segmented aortas with the strongest performance in zones 4, 5, and 9, providing a framework for further clinical applications. In our limited sample, there were no differences in baseline aortic zone volumes between rapid growth and no/slow growth patients.ARTICLE HIGHLIGHTSType of ResearchSingle-center retrospective cohort studyKey FindingsA deep learning model was developed to analyze volumetric growth in patients with medically managed acute uncomplicated type B aortic dissection. Volumetric growth was most pronounced in zones 5 (24%), 4 (13%), and 3 (11%). Model performance was best in zones 4, 5, and 9.Take Home MessageA trained, automated, open-source aortic zone segmentation model can accurately track changes in aortic growth by zone over time, providing framework for further clinical applications.Table of Contents SummaryA trained, automated model was developed to analyze aortic zone volumetric growth in a retrospective study of 59 patients with medically managed acute uncomplicated TBAD. Volumetric growth was most pronounced in zones 3-5, while model performance was best in zones 4,5, and 9.
Publisher
Cold Spring Harbor Laboratory