An Uncertainty-Distillation- and Voxel-Contrast-based Framework for One-shot Segmentation of Novel White Matter Tracts

Hao Xu, Tengfei Xue, Dongnan Liu, Fan Zhang, Carl-fredrik Westin, Ron Kikinis, Lauren Jean O'Donnell, Weidong Cai

Keywords: Diffusion MRI, White Matter Tract Segmentation, Uncertainty-Distillation, Contrastive Learning

Abstract: Diffusion-MRI-based white matter (WM) tract segmentation plays an important role in analyzing WM characteristics in healthy and diseased brains. The uncommon (novel) tract segmentation is important to the success of clinical brain operation and the reduction of postoperative complications. The massive WM tract annotations are time-consuming and need experienced neuroanatomists. Novel tract segmentation using only one annotated scan alleviates the above problems but is challenging. Existing fine-tuning-based studies achieve promising results but suffer from the feature overlap problem. In the work, we propose an uncertainty-distillation- and voxel-contrast-based one-shot novel WM tract segmentation framework, which includes an uncertainty distillation module to transfer semantic segmentation knowledge from base tracts to novel tracts and a voxel-wise multi-label contrastive module to adjust the feature embedding space so as to alleviate the feature overlap problem. We compare our method with several state-of-the-art (SOTA) methods that are designed to predict novel tract segmentation. The experimental results demonstrate that our method improves the one-shot segmentation accuracy of novel tracts in five experimental settings.

Submission Number: 39