Deep Learning for Region of Interest Based Clustering of White Matter Fibers

Feng-Sheng Tsai

Department of Biomedical Imaging and Radiological Science

China Medical University

fstsai@mail.cmu.edu.tw

    To cluster white matter fibers in whole-brain tractography, anatomical regions of interest (ROIs) are selected manually in brain diffusion MRI. Those ROIs are used to isolate tracts and cluster fiber bundles accordingly. Deep learning approaches may be applied to voxel-based ROI segmentation immediately; however, the number of voxels in ROIs is extremely smaller than the number of voxels in whole brain images, so they always suffer from the class imbalance problem when extracting related voxels of ROIs for training. Here we propose a hierarchical sampling technique to resolve the class imbalance problem of deep learning. ROI segmentation with deep learning is divided into hierarchical sub-tasks, from 2-dimensional objective-plane explorations to restricted, bounded hot-zone locations, and then to voxel-based discrimination. Sampling datasets in all sub-tasks are more balanced for training. Specifically, two ROIs for clustering arcuate fasciculus in whole-brain tractography are presented.