Automated Feature Segmentation of Ultra-Widefield OCT Images

To develop a lightweight neural network for automated cross-sectional and en face segmentation of ultra-widefield (UWF) OCT images acquired for retinopathy of prematurity screening.

Cross-sectional study.

Twenty-five infants with a birth weight <1500 g or gestational age <31 weeks were scanned using a portable, handheld, swept-source UWF-OCT device.

For cross-sectional B-scans, 3040 B-scans from 5 OCT volumetric scans obtained from 5 patients were segmented by 2 graders for the choroid and retina using custom-built tools in the Napari image viewer. Using these segmentations, a u-net with an EfficientNet-B0 backbone was trained in combination with task-specific augmentations to perform automated segmentation of the retina and choroid data with varying levels of image processing applied. For en face scans, 40 en face images from 20 unique patients were manually segmented by a single grader for retinal vessels. Using these segmentations, a u-net with an EfficientNet-B0 backbone was trained. Validation for both B-scans and en face images was performed using fivefold cross-validation. The fivefold cross-validation metrics were then compared with the metrics obtained by comparing grader segmentations.

The Dice similarity coefficient (DSC) was used to assess B-scan and en face segmentations.

The retinal and choroidal b-scan segmentations produced a DSC ± standard deviation of 0.925 ± 0.021 and 0.797 ± 0.062, respectively, averaged across the fivefolds. The en face vasculature segmentation produced a DSC ± standard deviation of 0.625 ± 0.0450.

Using u-net convolutional neural networks trained with task-specific augmentations, we developed en face and cross-sectional segmentations for UWF-OCT images, which will facilitate automated quantitative analysis with this novel modality.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.