Improving Aorta Segmentation from Phase Contrast MRI Using Adaptive Velocity-Dependent Weighting on the Deep Learning Output for Magnitude and Phase Images

Phase contrast MRI can provide a comprehensive analysis for the hemodynamic changes in the aorta which is useful for the diagnosis of several aortic diseases. However, an initial step of accurate segmentation of the aorta is necessary, which is usually a time-consuming and subjective step. Several methods have been proposed to automate this step using classical segmentation methods and recently deep learning models. Most of the current models combine the magnitude and phase images equally across all time phases which hinder the potential advantage that the frames of higher velocity might have more useful information compared to the low velocity frames. In this work, we propose a novel adaptive combination model that combines the output probability maps of both the magnitude and phase models based on an initial velocity estimation as a surrogate for the confidence level in the velocity images. We applied our model on the 2D-PC images of 215 patients and our results shows an accuracy of 87% for the magnitude images, 68% for the velocity images, 87.1% for the combined images, and 89.1 % for our proposed combination model.