Predicting Personalized Near-Normal Retinal Nerve Fiber Layer Thickness for Glaucoma Patients with Deep Learning

Abstract

Purpose : To develop a deep learning model to predict the personalized retinal nerve fiber layer (RNFL) thickness (RNFLT) map for glaucoma patients.

Methods : Reliable Cirrus optical coherence tomography (OCT) scans and 24-2 visual fields (VFs) were included in this study. First, we selected baseline OCT scans with 24-2 VFs tested within 30 days that were clinically normal meeting the criteria of mean deviation (MD) ≥ -1 dB, glaucoma hemifield test within normal limits and pattern standard deviation probability > 5%. Second, we selected follow-up OCT scans tested after the baseline OCT scans for each eye. Scanning laser ophthalmoscopy (SLO) funds photos were used to register follow-up RNFLT maps with corresponding baseline RNFLT maps. To avoid a trivial model to predict baseline RNFLT maps the same as the input follow-up RNFLTs as baseline and follow-up RNFLTs are highly similar, we developed a UNet-like (Figure 1 [a]) deep learning model using the follow up RNFLT map to predict the difference between baseline and follow-up RNFLT maps representing RNFL thinning. Our models were trained and tested using 2/3 and 1/3 of the entire dataset with patient level data separation. Performance was measured by the mean absolute error (MAE) and the Pearson correlation (R) between actual RNFL thinning map and predicted RNFL thinning map.

Results : 5,314 eyes from 5,314 patients were included in our study with average age and peripapillary RNFLTs of 60.8 ± 12.6 years and 85.9 ± 12.8 microns at baseline, respectively. 10,617 and 2,520 baseline-follow-up pairs of RNFLT maps were used for training and testing, respectively. The MAE and R score for the testing set were 10.8 and 0.45, respectively (Figure 1 [b]). Our model is able to predict RNFL thinning maps with a moderate accuracy. The first and second columns in Figure 2 show two examples of follow up RNFLT maps and actual versus predicted RNFL thinning maps. Our predicted RNFLT thinning maps were able to reproduce the actual RNFL thinning with low MAE errors (R: 0.76 and 0.79 for all pixels).

Conclusions : Our deep learning model can predict personalized near-normal RNFLT maps. Our model may help clinicians better detect RNFL defect in the early stage of glaucoma.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1

 

Figure 2