The Right Treatment for Abdominal Aortic Aneurysms: Machine Learning Based Predictions of Postoperative Endograft Apposition Using 4D Statistical Shape Models Including Thrombus

Determination of anatomic suitability is the most important factor in successful endovascular aneurysm repair (EVAR). As the function of the endograft is to prevent blood flow into the aneurysm sac from the healthy segment of the aorta, adequate seal is needed between proximal part of the graft and the aorta. Postoperative Shortest Apposition Length (SAL) less than 10mm following EVAR is associated with the development of a type 1a endoleak. Currently, procedures are planned by careful measurement of certain anatomic descriptors. The three dimensional morphology of the aorta is complex and might be oversimplified if only described by these anatomical parameters. In recent literature, a statistical shape model (SSM) is used to describe the morphology of the aortic neck. Also, a machine learning approach towards the preoperative prediction of SAL has been suggested. These recent approaches to predict postoperative SAL based on an SSM of the preoperative aortic neck used segmentations of the lumen of the aorta. It is hypothesized that for the prediction of postoperative SAL, also information about preoperative thrombus location, degree of circumference and thickness is needed. The aim of this study is to incorporate information of intramural thrombus into 4D SSMs and to provide insight in the possibility for continuous predictions of endograft apposition in the pararenal aorta. A dataset is composed of patients that received EVAR and patients that received FEVAR. Manually obtained wall segmentations of the aorta are used in the parameterization process before SSM creation. Resulting thrombus thickness between the lumen and wall segmentations is incorporated in the SSM as the fourth dimension. The SSM is validated on compactness, generalization ability and specificity. Ensemble learning, in the form of Random Forest classifiers, is used to predict postoperative SAL labels < 10 mm or ≥ 10 mm, by the means of leave-one-out cross validation. Predictions of classification are compared to previously published methods, using the DeLong’s test. Secondly, the possibility for continuous outcome of postoperative SAL is investigated in the pararenal aorta and the aortic neck specifically, by the use of regression algorithms. The root mean squared error (RMSE) is calculated for the continuous predictions and Bland-Altman plots are published to show potential bias.