Segmentation-Free OCT-Volume-Based Deep Learning Model Improves Point-Wise Visual Field Threshold Estimation

Abstract

Purpose : To overcome the floor effect of OCT measurements, we developed a deep learning model to estimate the functional deterioration directly from 3-dimensional (3D) OCT volumes (segmentation free) and compared the performance with the model trained with 2-dimensional (2D) OCT thickness maps (segmentation dependent) as inputs.

Methods : 8387 24-2 Humphrey visual field (VF) tests were acquired from 1129 patients over multiple visits. 13792 macular and 15026 optic nerve head scans (both 200x200 samplings, Cirrus HD-OCT, Zeiss, Dublin, CA) were acquired within 90 days of the VF visits. To reduce variances caused during imaging, the Bruch’s membrane surface was flattened and aligned by adjusting A-scans in the z-direction. Downsampling to 64x64x128 was applied to prevent memory shortage. Two models were trained to estimate the 52 VF threshold values: one was an 18-layer 3D ResNet taking OCT 3D volumes as inputs and the other was an 18-layer 2D ResNet taking 2D macular ganglion cell-inner plexiform layer and retinal nerve fiber layer thickness maps as inputs. A mean square error loss was used to train both models. The 52-point average of the mean absolute error (MAE) and Pearson correlation (PC) between the measured and estimated VF threshold values of each point were used to evaluate the performance.

Results : The MAE was significantly lower in 3D than 2D models (3.41 vs. 3.73 dB, respectively, p<0.001, Wilcoxon Signed-rank test). The PC was slightly better in 3D than 2D models (0.76 vs. 0.71, respectively, p<0.001, Williams test for equality of correlations). Figure 1 shows the histogram of VF threshold values in the training set with the MAE results. Both models performed better for VF threshold values between 20 and 35 dB, which were most frequently sampled in our dataset. For values under 20 dB, MAE of the 3D model clearly shows a better trend than that of the 2D model, indicating the model using 3D volumes may have less influence from the floor effects.

Conclusions : 3D model shows better accuracy than the 2D model, indicating the possibility of overcoming the floor effects of OCT measurements, although a relatively large error is present due to the under-represented low threshold sensitivity. Further investigation is needed with additional low threshold data.

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