Robust Optic Nerve Segmentation on Clinically Acquired CT.

Area of Science:

• Medical Imaging
• Neuroscience
• Ophthalmology

Background:

• Accurate segmentation of the optic nerve is vital for diagnosing neurological and ophthalmological conditions.
• Existing automated segmentation methods have limitations, particularly for smaller anatomical structures.
• Multi-atlas segmentation techniques show promise but require adaptation for precise optic nerve analysis.

Purpose of the Study:

• To develop and evaluate a robust, fully automated framework for segmenting optic nerves, eye globes, and muscles.
• To optimize registration and label fusion parameters for accurate segmentation of small central nervous system structures.
• To validate the framework's performance on diverse computed tomography (CT) datasets, including patient populations.

Main Methods:

• Utilized a robust registration procedure accounting for variable voxel resolution and field-of-view.
• Employed an optimal combination of SyN registration and the Non-local Spatial STAPLE label fusion algorithm.
• Validated the framework on 30 diverse human brain CT scans and 316 CT scans from thyroid eye disease (TED) patients.

Main Results:

• Achieved a median Dice similarity coefficient of 0.77 for optic nerve segmentation.
• Reported a median symmetric mean surface distance error of 0.55 mm and Hausdorff distance error of 3.33 mm for optic nerves.
• Demonstrated robustness by successfully segmenting optic nerves across a large patient cohort with thyroid eye disease.

Conclusions:

• The developed framework provides a robust and fully automated solution for optic nerve, eye globe, and muscle segmentation.
• The optimal registration and label fusion pipeline effectively addresses challenges in segmenting small anatomical structures.
• This automated approach has significant potential for clinical applications in diagnosing and monitoring diseases affecting the optic nerve.