Structured Learning for 3-D Perivascular Space Segmentation Using Vascular Features

Objective: The goal of this paper is to automatically segment perivascular spaces (PVSs) in brain from high-resolution 7T magnetic resonance (MR) images. Methods: We propose a structured-learning-based segmentation framework to extract the PVSs from high-resolution 7T MR images. Specifically, we integrate three types of vascular filter responses into a structured random forest for classifying voxels into two categories, i.e., PVS and background. In addition, we propose a novel entropy-based sampling strategy to extract informative samples in the background for training an explicit classification model. Since the vascular filters can extract various vascular features, even thin and low-contrast structures can be effectively extracted from noisy backgrounds. Moreover, continuous and smooth segmentation results can be obtained by utilizing patch-based structured labels. Results: The performance of our proposed method is evaluated on 19 subjects with 7T MR images, with the Dice similarity coefficient reaching 66%. Conclusion: The joint use of entropy-based sampling strategy, vascular features, and structured learning can improve the segmentation accuracy. Significance: Instead of manual annotation, our method provides an automatic way for PVS segmentation. Moreover, our method can be potentially used for other vascular structure segmentation because of its data-driven property.