During the past year, we continued a collaborative project with NEI Office of Scientific Information Officer (OISO) and BioTeam to develop automated segmentation of OCT images for retinal layers using machine-learning (ML) tools. We expanded our learning dataset to >1600 labelled images, and included both U-Net and Deeplabv3 model. The final product, AI-segmenter containing 12 independent models for segmentation of mouse OCT image. In addition, we also created a custom-made program to semi-automatically select the best segmented lines for the OCT images.
Using AI-segmenter to segment and measure different retinal layer thickness of wildtype (C57BL/6J) mice, we noticed inferior / superior asymmetry in the thickness of outer nuclear layer (ONL). ONL of inferior retina is a few micron thicker than superior retina, suggesting inferior contains more photoreceptor cells than superior retina does. On the other hand, the length from external limiting membrane (ELM) to retinal pigment epithelium (RPE) is longer for superior retina than the length for inferior. As ELM-RPE length is mainly determined by rod photoreceptor inner and outer segment length, the asymmetry of ELM-RPE length suggest there are more rod photoreceptor in superior retina than in inferior retina. Based on these observations, asymmetry of ONL thickness reflects the concentration of cone photoreceptors in inferior mouse retina.
To test our hypothesis, we examined RPE65 knockout mice and their littermate. It has been reported that cone photoreceptors degenerate first in this RPE65KO mouse model. We verified cone degeneration in 4-months old mice using whole-mount immunohistochemistry. For pigmented wildtype retina, there are more cone photoreceptors in inferior retina than those in superior retina. Degeneration of cone photoreceptors starts from inferior retina. Consequently, although both inferior and superior showed reduction in number of cone photoreceptors, inferior retina contains much less cones than superior retina for RPE65 KO mice. ONL thickness was measured from OCT images using AI-segmenter. For RPE65 KO mice, ONL thickness is thinner for inferior retina than the thickness of superior retina, whereas WT littermates showed thicker ONL for inferior retina than superior retina. These results support our hypothesis that asymmetry of ONL thickness is mediated by concentration of cone photoreceptors in the inferior mouse retina.
In conclusion, our study showed that AI-segmenter could provide accurate measure of retinal layer thickness for mouse OCT image, with sensitivity of micron level. Inferior / superior retina asymmetry detected on OCT image provides a non-invasive measure of rod and cone photoreceptor distribution in mouse retina.