Deep Learning-based Quantification of Anterior Segment OCT Parameters

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Anterior segment optical coherence tomography (ASOCT) is crucial for evaluating ocular structures.
• Accurate annotation of the scleral spur (SS) and segmentation of anterior chamber (AC) structures are vital for glaucoma diagnosis and management.
• Manual segmentation is time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop and validate a deep learning algorithm for automated SS annotation and AC structure segmentation in ASOCT scans.
• To enable precise measurements of AC, iris, and angle width parameters.
• To reduce subjectivity and improve efficiency in ASOCT-based ocular measurements.

Main Methods:

• Image contrast enhancement using CycleGAN.
• A heat map regression approach with a coarse-to-fine framework for SS annotation.
• An ensemble network (U-Net, full resolution residual network, full resolution U-Net) for structure segmentation.
• Comparison of algorithm-derived measurements with manual annotations (ground truth).

Main Results:

• The algorithm achieved high accuracy in SS annotation (Euclidean distance: 124.7 μm, ICC ≥ 0.95, error rate: 3.3%).
• Excellent segmentation performance with Dice similarity coefficient ≥ 0.91 for cornea, iris, and AC.
• Angle width measurements showed strong agreement with manual methods (≥ 95% within limits-of-agreement, ICC 0.71-0.87).
• Algorithm measurements demonstrated less variability compared to a semi-automated program.

Conclusions:

• A validated deep learning algorithm effectively automates SS annotation and AC structure segmentation in ASOCT scans.
• The algorithm performs comparably to human experts in both open-angle and angle-closure eyes.
• This technology significantly reduces measurement time and subjectivity, aiding clinical practice.