Staging Structural Damage in Glaucoma Based on Optical Coherence Tomography

Abstract

Purpose : To develop a glaucoma staging system based on optical coherence tomography (OCT)-derived retinal nerve fiber layer (RNFL) thickness measurements.

Methods : We developed a glaucoma damage classification system based on unsupervised k-means and Bayes minimum error classifier using 6561 RNFL profiles from 2269 eyes of 1171 subjects. We annotated the discovered clusters to different severity levels based on their respective mean global RNFL thickness. To establish an objective criterion for glaucoma staging, we computed optimal global RNFL thickness thresholds that discriminated different severity levels with highest accuracy using Bayes principle. We evaluated the quality of learning based on the area under the receiver operating characteristic curve (AUC).

Results : The k-means discovered four clusters with 1382, 1613, 1727, and 1839 samples and mean global RNFL thickness of 58.3 μm (±8.9:SD), 78.9 μm (±6.7), 87.7 μm (±8.2), and 101.5 μm (±7.9), respectively. The Bayes minimum error classifier identified optimal global RNFL thresholds of 70 μm, 85 μm, and 95 μm for discriminating the severity levels. The AUC was 0.90 and about 4% of normal eyes and 98% of eyes with advanced glaucoma had either global or quadrant RNFL thickness outside of normal limit.

Conclusions : Machine learning discovered optimal OCT-derived RNFL thresholds of about 70 μm, 85 μm, and 95 μm for staging glaucoma to normal, early, moderate, and advanced stages. The proposed staging system is unbiased with no pre-assumption or human expert intervention in the development process. Additionally, it is objective, easy-to-use, and consistent, which may augment glaucoma research and day-to-day clinical practice.

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

 

Figure 1. Diagram of the system. Circular RNFL thickness profiles (768 A-scans) were averaged to generate 64 sectors. These sectors and global quadrant measurements were input to the unsupervised learning model to identify clusters. Clusters were annotated based on their respective mean global RNFL thickness. Bayes classifier was used to identify optimal thresholds for staging severity.

 

Figure 2. Distribution of global RNFL thickness in four clusters identified by unsupervised k-means. Dashed lines represent the optimal thresholds for staging glaucoma. The proposed model suggested the global RNFL thickness thresholds of 70 μm, 85 μm, and 95 μm for staging glaucoma to normal, early, moderate, and advanced levels.