Diagnosis of central serous chorioretinopathy using deep learning with choroidal vascular en face images

Abstract

Purpose : To classify the central serous chorioretinopathy (CSC) using the deep learning (DL) with choroidal vascular en face images in optical coherence tomography (OCT).

Methods : One-hundred eyes were studied （53 eyes with CSC and 47 normal eyes）. Volume scans (12×12mm-square) were obtained at the same time as OCT angiographic scans (Plex Elite 9000 Swept-Source OCT®, Zeiss). High-quality choroidal vascular en face images of 53 eyes with CSC and 47 normal eyes at the segmentation slab of one-half of the subfoveal choroidal thickness (SCT) were created for analysis. Whole 100 en face images were split into 80 for training (100 times) and 20 for validation. Neural Network Console (NNC) developed by Sony and Keras backed up by Tensoflow developed by Google were used as the software for the classification with deep 16 layers of convolutional neural network.

Results : SCT was 480±92μm in CSC which was significantly thicker than 292±64μm in the normal eyes (P <0.01). In 20 eyes for validation including 8 eyes with CSC, the validation accuracy rate of NNC consisting was 100% (20/20) in NNC and 95% in Keras. There was no significantly difference.

Conclusions : Even if DLs with different programs, convolutional layer structures, and small data sets are used, CSC can be automatically classification with high accuracy from choroidal vascular en face images. DL can help in CSC diagnosis.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.