Abstract
Purpose :
Focal laser retinal injury models are a robust platform for in vivo testing of retinal regeneration strategies. However, quantitative assessment of the retinal structural changes resulting from laser photodamage is prohibitively labor-intensive, requiring manual segmentation of longitudinal volumetric datasets. Here, we present a deep-learning based approach for automated qualitative assessment of lesion volumes using OCT images to enable real-time assessment of injury severity and longitudinal tracking of tissue response to photodamage.
Methods :
In vivo OCT retinal imaging was performed longitudinally to track retinal injury response and analyze the reproducibility of retinal photodamage in murine models. Lesions with different severity were induced by varying the laser pulse duration, pulse power, and number of pulses. OCT images were denoised using self-fusion and fed into a convolutional neural network to segment lesion cross-sections. The network was trained to quantify photodamage between the outer plexiform layer (OPL) and retinal pigmented epithelium (RPE) accurately without the need for extensive image pre- and post-processing. The consensus of manually labeled OCT cross-sections from two independent image annotators were used as ground-truth.
Results :
The automated method was compared against the manual segmentation using the Dice coefficient achieving a score of 0.91 for training and 0.87 for testing. Figure 1 shows high degree of spatial co-registration between the manual (B) and automated lesion segmentation (C). Likewise, projections of lesion height of manually (E) and automatically (F) segmented lesions overlaid on the OCT en face projections demonstrate that both methods are well-correlated.
Conclusions :
Automated volumetric OCT segmentation enables efficient and robust quantification of lesion severity after laser delivery and during longitudinal follow-up. The proposed approach will allow for large-scale quantitative injury studies by eliminating the need of burdensome manual lesion segmentation. While promising, this initial proof-of-concept study will need to be validated on larger comparative studies with multi-grader segmentations that evaluate benefits in segmentation sensitivity and specificity.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.