Explainable Deep Learning Prediction of 10-2 Visual Field from 24-2 Visual Field Using Transformer

Abstract

Purpose : To develop a deep learning model to predict 10-2 visual field (VF) from 24-2 VF using a transformer.

Methods : We selected 10-2 VFs tested within 30 days of the corresponding 24-2 VFs from the Glaucoma Research Network. All VFs were analyzed in the right-eye format. Two types of 24-2 VF features were used (Figure 1 [a]): (1) 52 total deviation (TD) values; (2) 9 VF summary parameters including mean deviation (MD), pattern standard deviation (PSD), MD probability, foveal sensitivity, foveal sensitivity probability, laterality, test duration, time gap between 24-2 and 10-2 VFs, and age. We inputted the TD values into a multilayer perceptron (MLP) and the VF summary features into a transformer network, which were integrated to predict the 68 10-2 TD values. We used three-fold cross-validation with patient-level data separation. Performance was measured by the mean absolute error (MAE) and the Pearson correlation (R) between actual 10-2 VF and predicted 10-2 VF. The importance of different TD and VF summary features was determined by model weight analyses to explain our model.

Results : We included 5,189 24-2 reliable VF tests from 2,236 patients with an average age of 63.2 ± 15.2 years. The average MDs of the 24-2 and 10-2 VFs were -10.2 ± 8.7 dB and -4.9 ± 6.1 dB, respectively. The average MAE and R score for 68 10-2 VF test points were 3.4 ± 0.5 dB (ranging from 2.3 to 4.3 dB) and 0.84 ± 0.07 (ranging from 0.66 to 0.91), respectively (Figure 1 [b] and [c]). The prediction accuracy is lower in the inferior temporal region, which was previously reported as the less vulnerable zone by Hood and coworkers. As expected, 24-2 VF locations closer to the fixation were more important for the 10-2 VF prediction. However, 24-2 VF locations outside the central 10 degrees also moderately contribute to the 10-2 VF prediction (Figure 1 [d]). PSD and time gap were the two most important 24-2 VF summary parameters for 10-2 VF prediction. In two patient examples of 24-2 VFs and actual versus predicted 10-2 VFs (Figure 2), there was no clear evidence of central VF loss in the 24-2 VFs (Figure 2). Our predicted 10-2 VFs reproduced the superior temporal defect even though the MAE error was relatively high.

Conclusions : Our deep learning model can predict 10-2 VFs from 24-2 VFs. Our model may help clinicians decide when to additionally test patients with 10-2 VF apart from 24-2 VF.

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

 

Figure 1

 

Figure 2