Automated Detection of Posterior Vitreous Detachment on Optical Coherence Tomography Images Using Computer Vision and Deep Learning Algorithms

Abstract

Purpose : Recognizing a posterior vitreous detachment (PVD) is important for pre-surgical planning and risk-stratification for retinal tears and detachment. Deep learning (DL) methods of optical coherence tomography (OCT) images have been trialed on several retinal diseases to assist diagnosis; however, reliable models capable of detecting PVD have not been established. Our studies aim to develop reliable computer algorithms for automated detection of PVD on OCT images.

Methods : Consecutive OCT volumetric scans of 800 eyes obtained with Heidelberg Spectralis from November 2020 to October 2021 were retrospectively reviewed. Eyes with poor image quality or history of pars plana vitrectomy were excluded. Four trained graders individually labeled the PVD status of each eye by reviewing the entire volume B-scans, in addition to the horizontal and vertical raster scans. A second review was performed by an expert grader if consensus review was needed.

Two computer algorithms were developed. A customized computer vision algorithm based on image filtering and edge detection was designed to localize the posterior hyaloid and identify the presence or absence of a complete PVD in OCT volume scan. A second DL image classification model based on ResNet-50 architecture was trained to detect PVD status on OCT.

Results : Both the customized algorithm and DL model detection results were largely in agreement with the PVD status labeled by trained graders, with the customized algorithm surpassing the DL model in F1-scores for per-volume PVD detection. The accuracy of the customized algorithm and DL model was 85.57%, and 76.90%, respectively. The sensitivity of the customized algorithm and DL model was 78.21% and 86.09%, while the specificity was 93.62% and 70.45%, respectively (Table 1).

Conclusions : Using both traditional computer vision and deep learning approaches, we successfully developed reliable models to recognize PVD status on OCT images, demonstrating the potential for automated image classification to detect vitreoretinal pathologies and assist with pre-surgical planning. Further optimization of the two algorithms are currently in process with plans for validation using a prospective, independent data set.

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