Selective ensemble methods for deep learning segmentation of major vessels in invasive coronary angiography

Abstract

AbstractInvasive coronary angiography is a primary imaging modality that visualizes the lumen area of coronary arteries for the diagnosis of coronary artery diseases and guidance for interventional devices. In the current practice of quantitative coronary analysis (QCA), semi-automatic segmentation tools require labor-intensive and time-consuming manual correction; this limits their application in the catheterization room. For a more automated QCA, it is necessary to minimize operator intervention through robust segmentation methods with improved predictability. In this study, we introduced two selective ensemble methods that integrated the weighted ensemble approach with per-image quality estimation. In our selective ensemble methods, the segmentation outcomes from five base models with different loss functions were ranked by mask morphology or estimated dice similarity coefficient (DSC). The final output was determined by imposing different weights according to the ranking. The ranking criteria based on mask morphology were determined empirically to avoid frequent types of segmentation errors, whereas the estimation of DSCs was performed by comparing the pseudo-ground truth generated from a meta-learner. In the assessment with 7,426 frames from 2,924 patients, the selective ensemble methods improved segmentation performance with DSCs of up to 93.11% and provided a better delineation of lumen boundaries near the coronary lesion with local DSCs of up to 94.04%, outperforming all individual models and hard voting ensembles. The probability of mask disconnection at the most narrowed region could be minimized to <1%. The robustness of the proposed methods was evident in the external validation. Inference time for major vessel segmentation was approximately one-third, indicating that our selective ensemble methods may allow the real-time application of QCA-based diagnostic methods in routine clinical settings.