Synthesizing retinal degree of polarization uniformity from OCT with OCT-angiography by deep learning

Degree of polarization uniformity (DOPU) provides promising biomarkers of abnormalities in the retinal pigment epithelium (RPE) and is obtained by polarization-sensitive optical coherence tomography (OCT), which is not commercially available. A U-Net shape model was used to synthesize retinal DOPU from OCT with OCT angiography (OCTA) images. Sets of OCT, OCTA, and DOPU images from 175 subjects, 107 subjects, and 30 subjects were used for training, validation, and evaluation, respectively. RPE abnormalities were compared between True DOPU and synthesized DOPU. Healthy structure, RPE elevation, and RPE thickening were synthesized with high recall and precision. However, further improvements are required for RPE defect and hyperreflective retina foci synthesizing.

Automated summary

A U-Net shape model was used to synthesize retinal DOPU from OCT with OCTA images. The synthesized DOPU showed high recall and precision for healthy structure, RPE elevation, and RPE thickening. However, improvement is needed for synthesizing RPE defect and hyperreflective retina foci.