Multi-Task Learning based 3-Dimensional Striatal Segmentation of MRI - a Multi-modal Objective Assessment

The human striatum is a collection of subcortical nuclei that serves as a critical node in an information processing network. Recent studies have established a clear topographical and functional organization of projections to and from striatal regions that involves detailed and complex subdivisions of the striatum. Manual segmentation of these functional subdivisions is labor-intensive and time-consuming, yet automated methods have also been a challenging computational problem. Recently, deep learning has emerged as a powerful tool for various tasks, including semantic segmentation. Multi-Task Learning (MTL) is a machine learning technique that allows latent representations between related tasks to be shared during model prediction of multiple tasks with better accuracy. We utilized MTL to segment subregions of the striatum consisting of pre-commissural putamen (prePU), pre-commissural caudate (preCA), post-commissural putamen (postPU), post-commissural caudate (postCA), and ventral striatum (VST). Dice similarity coefficients (DSC) demonstrate excellent spatial agreement between manual and MTL-generated segmentations ([≥] 0.72 across all striatal subregions). Further quantitative task-based analysis was also conducted. Binding potential values, BPND, of [11C]raclopride PET, and ROI time-series and whole-brain connectivity using fMRI images were compared between results generated from manual segmentations and MTL-generated segmentations. BPND values from MTL-generated segmentations were shown to correlate well with manual segmentations with R2 [≥] 0.91 in all caudate and putamen subregions, and R2=0.69 in VST. Mean Pearson correlation coefficients of the fMRI data between MTL-generated and manual segmentations were also high in time-series ([≥]0.86) and whole-brain connectivity ([≥]0.89) across all subregions. We conclude that the MTL approach is a fast, robust and reliable method for 3D striatal subregion segmentation with results comparable to manually segmented ROIs.