Fully Automated Segmentation of Human Eyeball Using Three-Dimensional U-Net in T2 Magnetic Resonance Imaging

Area of Science:

• Ophthalmic imaging and analysis
• Medical artificial intelligence
• Deep learning in radiology

Background:

• Accurate segmentation of the human eyeball is crucial for ophthalmic diagnosis and research.
• Existing segmentation methods may lack automation or robustness across varying conditions.
• Deep learning models offer potential for automated and precise medical image analysis.

Purpose of the Study:

• To develop and validate a 3D U-Net deep learning tool for automated human eyeball segmentation.
• To compare the 3D U-Net's performance against a 2D U-Net and semiautomatic ground truth.
• To investigate age and sex influences on eyeball volume and gaze-dependent volume consistency.

Main Methods:

• Retrospective collection of 474 MRI scans from 119 patients with varying gaze directions.
• Application of 10-fold cross-validation for training and testing both 3D and 2D U-Net models.
• Quantitative performance metrics and Bland-Altman analysis for accuracy assessment; statistical analysis for age/sex and gaze effects.

Main Results:

• The 3D U-Net achieved high accuracy (>0.95) and acceptable agreement with ground truth, with slight overestimation.
• Significant age and sex differences in eyeball volume were detected by both methods.
• Segmentation accuracy varied by gaze direction, with upward gaze showing lower performance; no significant volume differences across gazes were found.

Conclusions:

• The 3D U-Net demonstrates effective and reliable automated segmentation of the human eyeball from T2-weighted MRI.
• The model's robustness across populations and gaze directions supports its utility in ophthalmic applications.
• This tool facilitates 3D eye globe shape analysis and eye movement detection in ophthalmic research.