Segmentation of Deep Gray Matter Nuclei with Imperfect Annotations and Anatomy A Priori Embedded in Template

Motivation: Automated segmentation enables objective and repeatable quantitative analysis of deep gray matter nuclei, which is essential to Parkinson’s disease (PD) studies. Goal(s): To combine the strengths of a classic segmentation algorithm and deep learning to achieve robust segmentation of deep gray matter nuclei without manual annotation. Approach: A brain nuclei template was created to generate template-based ROIs containing anatomical priori information. A classic segmentation algorithm was used to create imperfect algorithm-based ROIs, which were combined with template-based ROIs for training of a segmentation deep learning (DL) model. Results: The proposed model has achieved encouraging results, and still has room for improvement. Impact: Accurate and automatic segmentation for deep gray matter nuclei is essential to PD studies. The proposed DL segmentation model requires no manual annotations and may make the automatic segmentation more accessible for large datasets.