Pseudo-degree-of-polarization-uniformity image generated from OCT intensity using deep neural network

Abstract

Purpose : Polarization-sensitive OCT (PS-OCT) is expected to be the next generation OCT for clinical applications. The degree of polarization uniformity (DOPU) obtained by PS-OCT visualizes melanin distribution in retinal pigment epithelial cells (RPE) and depicts RPE abnormalities. However, the requirement of additional hardware for PS-OCT instrumentation has hindered the spread of PS-OCT. Here, we generated a pseudo-DOPU from ordinary OCT intensity using a convolutional neural network (CNN) and investigated its performance to detect abnormalities.

Methods : Fifteen eyes of 12 normal cases and 157 eyes of 123 abnormal cases were imaged by PS-OCT to obtain DOPU and OCT images. A CNN was trained to generate DOPU from OCT images using 10 eyes from 7 normal cases and 137 eyes from 103 cases of diseases.
Twenty eyes from 20 cases of disease and 5 eyes from 5 normal cases were used for evaluation. Ophthalmologist 1 extracted abnormal suprachoroidal regions from the OCT images, and ophthalmologist 2 identified RPE defects, thickening, prominences, and hyperintense retinal foci (HRF) in the DOPU and pseudo-DOPU faults at five locations within the regions. In normal eyes, the presence of abnormalities was assessed at five locations in each case.

Results : Representative pseudo-DOPU images of normal and abnormal cases which are not used in training are shown in Fig. 1. In these cases, pseudo-DOPU images well resemble DOPU images. In the disease case, RPE elevations (black arrows) and an RPE defect (white arrows) are in good agreement.
The recall and precision of pseudo-DOPU with respect to DOPU were as follows: for RPE deficiency (0.86, 0.37), for thickening (0.93, 0.65), for prominence (0.98, 0.98), and for HRF (0.74, 0.39). Of the 25 pseudo-DOPU cross-sectional images of normal eyes, 3 showed RPE defects, 1 showed RPE migration, and 1 showed abnormally low DOPU values, which were not confirmed by DOPU.

Conclusions : For the RPE abnormalities shown in DOPU, the pseudo DOPU did not show high precision but showed high sensitivity. For normal eyes, pseudo-DOPU seems to be similarly informative to PS-OCT, but we need to care about false abnormal findings.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

DOPU, and pseudo-DOPU B-scans of a normal and patient eyes.

 

Recall and precision of RPE abnormalities with generated pseudo-DOPU.