Using Artificial Intelligence to Screen for Retinal Tears in Acute, Symptomatic Posterior Vitreous Detachments

Abstract

Purpose : Approximately 15% of individuals with acute, symptomatic posterior vitreous detachments (PVD) will present with a retinal tear (RT). When PVDs are accompanied by visible pigment in the vitreous on fundus exam (Shafer sign), the risk of retinal tear rises to 60%. These pigment cells can be detected by spectral domain optical coherence tomography (SD-OCT) as isolated hyperreflective dots in the vitreous and are termed posterior vitreous opacities (PVO). Quantifying PVOs may serve as an adjunctive screening tool for RTs in patients with symptomatic PVDs. We performed a retrospective, observational clinical study to train a deep learning model to identify PVOs.

Methods : Patients presenting with acute, symptomatic PVD and concurrent SD-OCT imaging from Columbia University Irving Medical Center from 2011 to 2023 were studied. SD-OCT images were analyzed for patients presenting with flashes and/or floaters ≤ 32 days without prior history of uveitis, diabetic retinopathy, retinal vein occlusion, retinal tear, and vitrectomy. Only patients with high-definition 5-line raster scans taken on Cirrus^TM HD-OCT with a signal to noise ratio > 0.6 were included. Two independent reviewers annotated PVOs on SD-OCT images and a referee resolved disputes in PVO counts. A 2D U-net model was trained on the refereed images for automated localization of PVOs. SD-OCT images with PVOs were flagged for RT based on a PVO severity classification scheme (low, medium, and high PVO count) from Oh et al. (2017). The efficacy of automated PVO count was compared to other clinical methods of RT diagnosis using refereed images as the ground truth.

Results : 100 eyes in 93 patients (35 male, 65 female) met criteria for the study of which 10 eyes had a RT (10%). The ability for the U-net model to detect RTs by PVO count performed comparably to other methods of diagnosing RTs (table 1). Additionally, the U-net model had a range of difference between predicted and actual PVO count of [-9.6, 8.8] with a mean difference of -0.36 ±1.96 (figure 1).

Conclusions : 2D U-Net learning architectures can reliably localize PVOs on an OCT image and achieve similar, or even better, sensitivities and specificities for RT diagnosis compared to other methods of diagnosis. In combination with other diagnostic features of RTs, deep learning models can help clinicians evaluate patients with an acute, symptomatic PVD for a RT.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Table 1

 

Figure 1