Estimating severity of visual field loss from optical coherence tomography imaging using artificial intelligence

Abstract

Purpose : To assess the accuracy and reproducibility of a machine learning algorithm in estimating visual field severity from optical coherence tomography (OCT) data.

Methods : We developed an artificial neural network (ANN) to estimate visual field defect severity (mean deviation; MD) of the 24-2 tests (Humphrey visual field analyzer) from OCT-derived retinal nerve fiber layer (RNFL) thickness measurements (Circle scans, Spectralis). We used 1769 visual field and OCT pairs from 1769 eyes of normal subjects and glaucoma patients to train and test the ANN model. We employed two independent subsets consisting of 698 visual fields and OCT pairs from 698 eyes (Rotterdam Eye Hospital) and 256 visual fields and OCT pairs from 256 eyes (UCLA), to validate models and assess the reproducibility. Scatter plots of the estimation parameters across glaucoma severity were visualized to assess the models subjectively. Mean absolute error (MAE), root mean square error (RMSE), median of absolute error, and Pearson R metrics were computed to evaluate the accuracy of prediction objectively.

Results : The RMSE and MAE of the testing dataset were 5.2 dB and 4.0 dB, respectively. The RMSE and MAE of the models on an independent dataset from Rotterdam Eye Institute were 4.4 dB and 3.3 dB, respectively. The RMSE and MAE of the models on an independent dataset from UCLA were 5.3 dB and 3.9 dB, respectively (Table 1).

Conclusions : The proposed ANN model estimated MD with a reasonable accuracy. Contrary to most of the models in the literature, this model has a reasonable generalization to unseen data from other races. A successful development of such learning models may assist clinicians in estimating subjective visual field tests loss from objective OCT measures.

This is a 2021 Imaging in the Eye Conference abstract.

 

Table 1. Accuracy of the model in estimating visual field mean deviation (MD) from retinal nerve fiber layer (RNFL) thickness measurements.

 

Figure 2. Scatter plots of the true versus estimated mean deviations (MD) of left) Testing dataset, middle) Rotterdam dataset, and right) UCLA dataset.