Automated learning of glaucomatous visual fields from OCT images using a comprehensive, segmentation-free 3D convolutional neural network model

A segmentation-free 3D Convolutional Neural Network (3DCNN) model was adopted to estimate Visual Field (VF) in glaucoma cases using Optical Coherence Tomography (OCT) images. This study, conducted at a university hospital, included 6335 participants (12,325 eyes). Two models were trained, one on the Glaucoma-Specific Training Group (GTG) and one on the Comprehensive Training Group (CTG) that included various ocular conditions without manual preselection. The CTG showed significantly better performance than the GTG in estimating VF thresholds and Mean Deviation (MD) for both Humphrey Field Analyzer (HFA) 24-2 and HFA10-2 test patterns (p < 0.001). Strong correlations were observed between the estimated and actual VF thresholds for HFA24-2 (Pearson's r: 0.878) and HFA10-2 (r: 0.903), as well as MD for HFA24-2 (r: 0.911) and HFA10-2 (r: 0.944) in the CTG. The CTG demonstrated lower estimation errors than the GTG and smaller errors in severe cases. The model's performance remained relatively stable even in advanced glaucoma cases. The model's ability to learn from a comprehensive dataset without human annotation highlights its potential for large-scale training in the future, potentially improving glaucoma assessment and monitoring in clinical practice. Further validation in external datasets and exploration in different clinical settings are warranted.