Impact of training data composition on the generalizability of convolutional neural network aortic cross-section segmentation in four-dimensional magnetic resonance flow imaging

• DOI: 10.1016/j.jocmr.2024.101081
• Publication Date: 2024-08-08T17:43:44Z
• Authors: Chiara Manini; Markus Hüllebrand; Lars Walczak; Sarah Nordmeyer; Lina Jarmatz; Titus Kuehne; Heiko Stern; Christian Meierhofer; Andreas Harloff; Jennifer Erley; Sebastian Kelle; Peter Bannas; Ralf Felix Trauze...; Jeanette Schulz-M...; Anja Hennemuth
• Keywords: 4D flow MRI; Segmentation; Cardiovascular and Metabolic Diseases; RC666-701; Integrative Biomedicine [Topic 3]; Diseases of the circulatory (Cardiovascular) system; Deep learning; BAV; Thoracic aorta; Original Research

Time-resolved, three-dimensional phase-contrast magnetic resonance imaging (4D flow MRI) plays an important role in assessing cardiovascular diseases. However, the manual or semi-automatic segmentation of aortic vessel boundaries 4D data introduces variability and limits reproducibility hemodynamics visualization quantitative flow-related parameter computation. This paper explores potential deep learning to improve MRI by developing models for automatic analyzes impact training on generalization model across different sites, scanner vendors, sequences, pathologies. The study population consists 260 datasets, including subjects without known pathology, healthy volunteers, patients with bicuspid valve (BAV) examined at hospitals. dataset was split train subsets representations characteristics such as gender, age, model, vendor, field strength. An enhanced 3D U-net convolutional neural network (CNN) architecture residual units trained 2D+t cross-sectional segmentation. performance evaluated using Dice score, Hausdorff distance, average symmetric surface distance test data, datasets not represented set (model-specific), overall evaluation set. Standard diagnostic parameters were computed compared results Bland-Altman analysis interclass correlation. representation technical factors vendor strength had strongest influence performance. Age a greater than gender. Models solely BAV patients' performed well but vice versa. highlights importance considering heterogeneous widely applicable CNN segmentations MRI, particular focus inclusion pathologies aspects acquisition.