Systematic review finds variable MRI segmentation accuracy for pituitary adenomas and gland

This systematic review and meta-analysis examined 34 studies (from 353 reviewed) evaluating automatic and semi-automatic magnetic resonance imaging (MRI) segmentation methods for pituitary adenomas and the pituitary gland in human subjects. The majority of studies employed deep learning, U-Net-based models. No explicit comparator arm was defined; instead, performance was assessed using Dice overlap scores, a metric of segmentation accuracy.

For automatic segmentation methods, performance varied substantially: Dice scores ranged from 0% to 89% for pituitary gland segmentation and from 4% to 96% for adenoma segmentation. Semi-automatic methods demonstrated more consistent performance, with Dice scores of 80% to 92% for the pituitary gland and 75% to 88% for adenomas. The review did not report specific human subject numbers, adverse events, or tolerability data related to the imaging techniques.

Key limitations significantly constrain interpretation. Most studies failed to report critical methodological and clinical details: MRI field strength, patient age, adenoma size (macro/micro/giant), adenoma type, and exact human subject numbers were largely unreported. These omissions make it difficult to assess generalizability and compare studies directly.

The authors conclude that while automated segmentation techniques show promise—particularly for adenoma segmentation—further improvements are needed to achieve consistently good performance, especially for small structures like the normal pituitary gland. Future progress will require both methodological innovation and larger, more diverse datasets to enhance clinical applicability. For now, these tools remain primarily in the research domain.