Improving the detection of visual field progression in glaucoma using fused data from visual field testing and optical coherence tomography

Abstract

Purpose : Comprehensive evaluation of functional and structural changes is essential for detecting glaucoma progression. This study aims to develop and validate a model integrating data from visual field (VF) testing and optical coherence tomography (OCT).

Methods : We developed an autoencoder (AE) model to fuse pointwise contrast sensitivity data from VFs and retinal nerve fiber layer thickness (RNFLT) profiles from OCT. The AE model is designed to discover a compact encoding of the input VF data and the RNFLT profile data, with the encoding serving as the fused data. To train the model, we introduced an encoding loss to penalize the disparity between the encoding and the input VF data. This approach ensures that the AE-fused data maintains morphological similarity to the measured VF while retaining sufficient information for reconstruction.
For model evaluation, we labeled eyes as progressing if the linear regression slopes of the VFs’ mean deviation were worse than −0.5 dB/year, calculated from all measured VF data for each eye. Then, we used data from only the first 3 years to classify whether an eye was progressing. The classification sensitivity and specificity obtained from the AE-fused data were compared to those from the measured VF data. The performance of our AE model was also compared to a state-of-the-art Bayesian linear regression (BLR) model that integrates structural and functional data.

Results : A total of 2504 VF-OCT test pairs from 253 eyes of 140 glaucoma patients followed over an average of 7.7 years were evaluated. Using data from the initial 3 years (Figure 1), the AE-fused data achieved a mean classification specificity of 0.65. An 81% and 27% improvement over the specificity obtained from measured VF data (0.36) and the BLR model (0.51), respectively. Over the same period, AE-fused data demonstrated a mean sensitivity of 0.52, outperforming the BLR model (0.35) by 49% and slightly lower than the measured VF data (0.56). These advantages persisted whether a relaxed (-0.2 dB/year) or a strict (-1.0 dB/year) criterion for VF progression was applied.

Conclusions : An AE model that fuses structure and function data for the prediction of VF progression can improve the specificity while maintaining the similar sensitivity of methods that only use VF measurements.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Comparisons of the performance in detecting visual field progression