Deep Learning-Based Automated Detection and Quantification of Posterior Segment Inflammatory Features on SD-OCT

Abstract

Purpose : Objective characterization of posterior segment inflammation remains a challenging issue in disease management. Automated detection and quantification of posterior inflammatory signs could dramatically improve consistency in disease activity assessment. The purpose of this analysis was to evaluate the ability of a deep-learning segmentation platform to detect differences in posterior vitreous debris and preretinal inflammatory deposits between eyes with significant posterior uveitis and control eyes.

Methods : Initial model training on 400 OCT images with variable posterior segment inflammation was performed after annotation for cellular debris (i.e., preretinal deposits, vitreous cell). The segmentation platform characterized the number and relative location (preretinal vs vitreous). Following model development, a case-control independent validation assessment was performed on 21 eyes with posterior uveitis and 21 control eyes.

Results : The automated segmentation platform successfully identified both vitreous and preretinal debris (Figure 1). The total number of hyperreflective spots in the inflammation group was significantly higher than in the control group (2.33 ± 1.82 vs 0.50 ± 0.51, p <0.01). The number of preretinal deposits were significantly higher in the inflammation group (2.04 ± 1.75 vs 0.42 ± 0.41, p< 0.01). The number of hyperreflective spots in the vitreous was also significantly higher in inflammation group compared to the control (0.47 ± 0.45 vs 0.11 ± 0.32; p= 0.04).

Conclusions : These preliminary results demonstrate the feasibility of automated detection and objective quantification of inflammatory features in posterior uveitis from SD-OCT images. Future research will include further validation in larger dataset and longitudinal assessment of quantitative alterations following therapeutic interventions.

This is a 2021 Imaging in the Eye Conference abstract.

 

Fig. 1: Sample segmentation of horizontal optical coherence tomography B-scan image with inflammation, a, c) the original images b, d) the segmented overlay with identification of preretinal deposits (red), and vitreous debris (blue).