Generating novel pituitary datasets from open-source imaging data and deep volumetric segmentation.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Pituitary adenoma diagnosis is challenging due to complex imaging features and equivocal clinical signs.
• Existing whole-brain MRI datasets lack dedicated pituitary imaging for AI model development.
• Computer vision models offer potential for improved pituitary adenoma diagnosis.

Purpose of the Study:

• To develop a novel, large-scale dataset of pituitary imaging using deep volumetric segmentation.
• To leverage existing whole-brain MRI scans for pituitary region extraction.
• To address the challenge of limited pituitary imaging datasets for AI research.

Main Methods:

• Trained deep learning volumetric segmentation models on 318 manually annotated MRI scans.
• Utilized six open-source whole-brain MRI datasets for model development and testing.
• Validated model performance on 418 held-out MRIs from five different datasets.

Main Results:

• Achieved high agreement between manual and model volumetric segmentations (Dice scores 0.76-0.82).
• Successfully generated the first and largest pituitary imaging dataset, comprising 6,755 MRIs from six sources.
• Demonstrated robust out-of-distribution performance across diverse MRI data.

Conclusions:

• The developed deep volumetric segmentation models effectively generate pituitary imaging datasets.
• The resulting dataset is the largest to date and generalizes well across populations and scanner types.
• This dataset is poised to be a valuable resource for future machine learning research in pituitary pathologies.