Semi-supervised learning in perivascular space segmentation using MRI images

Yaqiong Chai, Hedong Zhang, Gilsoon Park, Erika Lopez, Cong Zang, Jongmok Ha, Omar Elhawary, Hosung Kim

Keywords: semi-supervised learning, perivascular space, MRI

Abstract: Accurate segmentation of perivascular space (PVS) is essential for its quantitative analysis and clinical applications. Various segmentation methods have been proposed, but semi-supervised learning methods have never been attempted. Here, a 3D multi-channel, multi-scale semi-supervised PVS segmentation (M2SS-PVS) network is proposed. The proposed network incorporated multi-scale image features in the encoder and applied a few strategies to mitigate class imbalance issue. The proposed M2SS-PVS network segmented PVS with the highest accuracy and high sensitivity among all the tested supervised and semi-supervised methods.