Motivation: Observer influence hampers reproducibility in quantitative analysis. However, there have been relatively few efforts toward creating automated analysis tools for carotid arteries. Goal(s): We aim to investigate and evaluate performance of AI-driven pipeline analysis to automatically segment carotids and quantify hemodynamics from 4D flow MRI. Approach: A dense U-net was trained manually-generated labels (n=265) to segment carotids in 4D flow MRI. Algorithmic analysis quantified three hemodynamic parameters. Segmentation similarity and parameter agreement was evaluated (separate group, n=66). Results: Good segmentation similarity was achieved (median DSC=0.88,0.89 left,right). Close limits of agreement for quantifications were observed, although there was slight (<5%) bias in some parameters. Impact: This study demonstrated a method for automatically generating segmentations of the carotid vessels in 4D flow MR images and quantifying hemodynamics. It can inform efforts to reproducibly study carotid hemodynamics in large cohorts of subjects.
Johnson et al. (Tue,) studied this question.