Comparative Assessment of Fully-Automated Machine Learning Enabled Fluid Feature Extraction to Expert-Reader Interpreted Fluid Segmentation in Diabetic Macular Edema and the Importance of Image Quality Stratification

Abstract

Purpose : The optimization of deep learning models for fluid feature extraction to enable reader-independent assessment of fluid presence and characteristics in retinal disease is an important milestone for automated image processing. This analysis aims to compare a next generation fully automated deep learning model for fluid segmentation against human generated output in diabetic macular edema (DME).

Methods : This was a retrospective assessment of eyes with DME undergoing treatment with anti-VEGF therapy. All eyes had concurrent SD-OCT scans. All scans were initially analyzed with an earlier generation intraretinal fluid (IRF) segmentation system with subsequent expert reader manual correction, as needed. A next generation fluid segmentation model with enhanced sensitivity and specificity was evaluated with automated volumetric outputs and compared to reader-edited findings. An automated SD-OCT quality assessment system was used to stratify scan quality as “good”, “moderate”, and “poor” for image quality-based assessment of performance.

Results : A total of 1572 OCT scans were included in this analysis.The model correctly identified the presence or absence of IRF in 96% of the scans. Scans identified as “good” quality demonstrated significantly closer volumetric agreement to expert readers than “moderate” and “poor” quality scans (p<0.001).

Conclusions : This fully automated fluid model demonstrated overall strong agreement with human edited fluid segmentation. Image quality appears to have an important impact on agreement between automated and manually-edited results. Future assessments will include using the next generation system as foundational starting-point for expert reader corrections for future comparisons and assessing minimum quality standards for automated image analysis platforms.

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