AI-based Clinical Assessment of Optic Nerve Head Robustness from 3D Optical Coherence Tomography Imaging

Abstract

Purpose : To develop a clinically applicable approach to assess the robustness of an individual optic nerve head (ONH) from a standard 3D optical coherence tomography (OCT) scan by predicting how it would deform under a hypothetical acute change in intraocular pressure (IOP).

Methods : 316 subjects had their ONHs imaged non-invasively with 3D optical coherence tomography (OCT) before and after acute IOP elevation through ophtalmo-dynanometry – a method to raise IOP via globe indentation. We then categorized each ONH as robust or compliant. To do so, a 3D digital volume correlation algorithm was applied to both OCT volumes of each ONH to extract an IOP-induced average effective strain in the lamina cribrosa region (E_eff). 169 subjects were considered to exhibit robust ONHs (E_eff < 4%) and 147 were classified as compliant (E_eff > 4%). Learning from these biomechanical data, we compared three algorithms to assess ONH robustness strictly from a baseline (undeformed) OCT volume: (1) a point cloud classification network, (2) a decision tree ensemble, and (3) an autoencoder in combination with a fully connected binary classification network. For each method, 253 subjects were used for training and 63 for testing. To evaluate the performance of our algorithms in predicting ONH robustness, we used the area under the receiver operating characteristic curve (AUC).

Results : All three methods were able to assess ONH robustness from 3D structural information alone. The point cloud classification network (AUC: 0.77) performed slightly better than the decision tree ensemble (AUC: 0.69) and the autoencoder (AUC: 0.68).

Conclusions : We introduce a novel machine learning based method to assess ONH robustness strictly from a standard 3D OCT scan. Our proposed approach may have wide clinical interest because it does not require new hardware (it can be combined with any existing OCT device) and can assess ONH robustness without the need of performing complex biomechanical tests. Longitudinal studies should establish if ONH robustness estimated by the herein presented technique helps to predict and better understand the development and progression of glaucoma. Our method will soon be improved by incorporating a larger biomechanical dataset.

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