Deep learning-based segmentation of OCT images for choroidal thickness.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate measurement of choroidal thickness is crucial for diagnosing and monitoring various eye conditions.
• Manual segmentation of optical coherence tomography (OCT) scans is time-consuming and subjective.

Purpose of the Study:

• To develop and validate a custom deep learning-based automated segmentation tool for choroidal thickness in OCT scans.
• To compare the performance of the in-house automated method against manual segmentation and an open-source algorithm.

Main Methods:

• A Deeplabv3+ network (ResNet50) was trained on 10,798 manually segmented OCT scans.
• Validation involved comparing manual and in-house automated segmentation on 130 unique scans using Bland-Altman analysis, ICC, and Deming regression.
• The in-house method was also benchmarked against an open-source algorithm.

Main Results:

• The in-house automated method showed no significant difference in mean choroidal thickness compared to manual segmentation across different regions (P > 0.05).
• Excellent agreement was observed between manual and in-house automated methods (ICC: 0.96-0.98, P < 0.001).
• An open-source algorithm yielded consistently thinner choroidal thickness measurements than both manual and in-house automated methods.

Conclusions:

• Custom deep learning automated choroid segmentation demonstrates excellent agreement with manual segmentation.
• The automated approach provides objective and efficient estimation of choroidal thickness.
• This technology has the potential to improve clinical workflows and diagnostic accuracy in ophthalmology.