A Multi-Stage Machine Learning Approach for Predicting Indicators of EVAR Stent Complications in CT Images - A Pilot Study on AI-Based Diagnosis of Post-EVAR Loss of Seal

Abstract

Abstract Patients that have received endovascular aneurysm repair (EVAR) surgery regularly undergo computed tomography (CT) scans to monitor the placement of the inserted stent. The process of analyzing these 3D volumes is, however, difficult and requires specialized expertise. To improve the detection of complications, new clinical procedure has been developed to diagnose loss of seal, a common complication of EVAR surgery. The protocol consists of three steps: centerline definition, identification of stent ends, and measurement of the seal zone length. This project aims to employ a machine learning (ML) based approach to CT analysis in order to diagnose loss of seal, with annotations from the described protocol. The aim is achieved by predicting the seal zone length through a novel sequential approach, and a baseline. The sequential approach consists of two models in sequence, where the intermediate output is the stent endpoint locations. In order to use ML methods, a dataset of 143 patients and a total of 399 CTs was curated from unprocessed clinical data. The results indicate that detecting the three stent endpoints with a transformerbased model is a promising first step towards diagnosis and could potentially automate one step in the clinical protocol. Regression of the seal zone length was, on the other hand, unsuccessful with the current model architecture and it was deemed a far more complex task. Several improvements could be made as future work, such as utilizing a transformer-based model for seal zone regression and predicting additional intermediate labels. The sequential approach has potential, but some steps could be tweaked to reach an accurate and generalizable method.