Deep Learning Approaches Can Detect Glaucomatous Functional Loss better than standard SD-OCT Retinal Nerve Fiber Layer Thickness

Abstract

Purpose : To develop and evaluate deep learning approaches to predict functional progression from spectral domain optical coherence tomography (SD-OCT) retinal nerve fiber layer thickness (RNFLt) measurements in primary open angle glaucoma (POAG).

Methods : A cohort of 375 subjects (538 eyes) with glaucomatous visual field damage (baseline mean MD=-4.5 dB) and 254 subjects (444 eyes) without visual field damage was followed longitudinally with twice per year visual fields (24-2 standard automated perimetry) and SD-OCT imaging (optic nerve head cube and circle scans). Images were divided into those collected from eyes with glaucomatous visual fields (three or more consecutive abnormal visual field results) and those without. Participants were randomly divided into a training set (90%, 566 subjects) and an independent testing set (10%, 63 subjects). Scanning laser ophthalmoscopy (SLO) images and RNFLt maps were extracted from each SD-OCT cube. RNFLt maps were extracted using the San Diego Automated Layer Segmentation Algorithm (SALSA). SLO and RNFLt images were used to train deep learning models (AlexNet) to identify those exhibiting glaucomatous visual field damage. Transfer learning was used by adopting models previously trained on a large, general image dataset (ImageNet). The models were fine-tuned on the SD-OCT training set and evaluated on the test set. A model trained only on SLO images, a model trained only on RNFLt maps, and a combined model were evaluated. For comparison, mean RNFLt measurements from circle scans were also evaluated. Models were compared using area under receiver operating characteristic curves (AUC).

Results : For detecting visual field damage from SD-OCT images, the combined SLO and RNFLt deep learning model achieved the highest AUC (0.92) among the all deep learning models. This model was significantly better than models built exclusively on either SLO (AUC=0.81, p<0.001) or RNFLt (AUC=0.85, p<0.001). It also significantly outperformed the circle scan mean RNFLt measurements (AUC=0.73, p<0.001).

Conclusions : Deep learning models can improve our ability to identify eyes with glaucomatous visual field damage based solely on structural SD-OCT imaging. The results suggest that deep learning models may be used to automate detection of POAG and provide better guidance for clinicians in identifying POAG than standard RNFLt measurements.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.