Leaflet thrombosis automatic identification in transcatheter aortic valves using 4DCT

Abstract

Leaflet thrombosis (LT) is a significant complication of transcatheter aortic valve implantation (TAVI) that impacts patient outcomes and transcatheter heart valves (THVs) long-term durability. Subclinical LT (SLT), manifested as hypo-attenuated leaflet thickening (HALT) and potential reduced leaflet motion (RELM), is challenging to diagnose due to its reliance on manual evaluation and observer variability. Although computed tomography (CT) is the preferred imaging modality for LT detection, manual assessment remains labor-intensive and prone to inconsistencies. Additionally, as TAVI procedures are increasingly performed in younger patients, concerns about THVs long-term durability —and particularly regarding LT— are growing. This study aims to develop automated segmentation models using the nnU-Net architecture to detect and characterize thrombi in 4DCT scans of TAVI patients. The methodology includes three main steps: manual annotation of the dataset, thrombus segmentation using eight distinct nnU-Net models, and evaluation based on segmentation metrics and clinical thrombus information. Several models achieved precision values exceeding 0.8 for LT patients, demonstrating the potential of automated segmentation to enhance LT detection. Furthermore, the observed variations in thrombus volume across the cardiac cycle highlight the importance of selecting the optimal phase for LT assessment, suggesting that dynamic evaluation could improve diagnostic accuracy. This work lays the groundwork for early LT detection and the development of predictive biomarkers, offering automated LT detection and characterization that reduces manual effort and observer variability. The proposed dynamic 4DCT analyses could improve LT diagnostics and inform personalized anticoagulation strategies, potentially leading to better long-term outcomes.