Dense Optic Nerve Head Deformation Estimated using CNN as a Structural Biomarker of Glaucoma Progression

Abstract

Purpose : Glaucomatous structural changes are observable non-invasively using optical imaging techniques. The aim of this study is to test the hypothesis that our newly designed convolutional neural networks (CNN) trained for estimating transformations in generic image sequences can be useful for estimating glaucomatous progression from confocal image sequences of the optic nerve head (ONH).

Methods : Optical flow provides a dense pixel-wise estimate of a scene transformation from an image sequence. We propose the average magnitude of flow velocities within the ONH region as a structural biomarker of glaucoma progression. Dense ONH structural transformation between a baseline scan and the most recent follow-up scan was estimated using various pre-trained CNNs namely FlowNet Correlation and FlowNet2¹, Teney², and our newly designed compact deep-dilated CNN (DDCNet) framework³. Diagnostic accuracy of the proposed flow-based structural biomarker was evaluated using longitudinal HRT-II exams of study participants in the UCSD Diagnostic Innovations in Glaucoma Study. 36 eyes progressed by stereophotos or visual field guided progression analysis, and 21 eyes were longitudinal normal eyes. Diagnostic sensitivity and specificity were estimated using a maximum likelihood classifier.

Results : The area under the ROC curve, sensitivity, and specificity of our deep-learning methods for estimating ONH structural transformations and other legacy computational methods (POD framework and TCA⁴) are presented in Table 1.

Conclusions : CNNs trained to estimate dense transformations in generic scenes were able to detect glaucomatous progression from HRT scans of the ONH with higher diagnostic accuracy. We anticipate that fine-tuning these networks using ONH sequences may further improve their diagnostic performance. Further, we conclude that the average magnitude of ONH structural changes estimated using the CNNs is a robust candidate biomarker for detecting glaucoma progression.

This is a 2021 ARVO Annual Meeting abstract.

 

 

Figure 1. ONH deformation field of a longitudinal normal eye (a) and a progressing eye (b) estimated using our DDCNet optical flow CNN model. For clarity in illustration, deformation field within the optic disc region is magnified. In addition to quiver plot, deformation field is also displayed using the color coding scheme (c) to depict the direction and magnitude of tissue remodeling.