Automated Segmentation of the Pituitary and Pineal Glands.

Abstract

The pituitary and pineal glands are two small yet critical brain structures that help to modulate the human endocrine system. Unfortunately, very little research has been devoted to segmenting the pineal gland, and existing methods for pituitary segmentation focus only on the entire gland without distinguishing between its two lobes. To fill this gap, this work presents the first deep-learning-based tool for segmentation of both the pineal and pituitary glands in T1-weighted MRI. A five-fold cross-validation study was conducted on a manually labeled training dataset and produced segmentations with accuracy comparable to similar methods for segmenting other small brain structures. Model performance was then tested in three publicly available datasets using a total of n = 816 subjects, the results of which were both highly reproducible and robust to differences in MRI scanners and acquisition protocols. Finally, an analysis was performed to identify group differences related to sex and the diagnosis of schizophrenia and showed that volumes measured from the output segmentations were effective at discerning sex- and disease case-related differences in the pituitary and pineal glands.