Clinical Evaluation of nn-UNet for Automated Segmentation of Pituitary Gland and Optic Apparatus in Brain MRI: A Multi-Database Approach

Motivation: Automated segmentation of the pituitary gland, stalk, and optic apparatus in brain MRI is essential for enhancing diagnostic accuracy in evaluating sellar region disorders. Goal(s): To develop and assess nn-UNet-based deep learning models for segmenting sellar region in MRI images. We seek to determine their performance across three datasets, including clinically acquired images. Approach: We trained the nn-UNet models on research datasets from the Hammers and MRXFDG databases and evaluated them using a separate clinical dataset. Results: The nn-UNet models demonstrated variable performance, achieving optimal results when tested on similar datasets. The Unified-Trained model showed promise, but the clinical data revealed significant limitations. Impact: This research has the potential to improve the accuracy of MRI diagnostics for pituitary and sellar region disorders. By addressing the challenges of model performance on clinical data, it opens new avenues for optimizing deep learning applications in medical imaging.