Abstract
Purpose :
We have already developed a rapid screening for glaucoma using binocular perimetry Imo (Imo screening program; ISP). Predicting visual field (VF) parameters such as patterns or ratios of visual loss would provide more crucial functional information about the severity of glaucoma. The aim of this study is to simulate VF parameters of the widely used Humphrey Visual Field Analyzer (HFA) using ISP data, thereby allowing more precise clinical decisions even in screening settings.
Methods :
We retrospectively enrolled patients undergoing VF measurements with ISP and HFA 24-2 on the same day in Jikei University Hospital, excluding those with reliability lower than criteria or with diseases affecting VF other than glaucoma. We developed a multitask neural network to predict VF parameters of HFA 24-2: probability plots of total deviation (TD) and pattern deviation (PD), mean deviation (MD), pattern standard deviation (PSD), visual field index (VFI), and glaucoma hemifield test (GHT). We also evaluated the efficacy of pre-training and zero-shot learning using an augmented ISP dataset, created by combining 20 points from HFA 24-2 and 8 points from HFA 10-2 conducted within 6-months, with thresholding applied to these 28 points. Model performance was evaluated based on F1-score and mean absolute error (MAE).
Results :
The actual ISP dataset included 187 eyes of 112 patients, and the augmented ISP dataset had 3470 eyes of 883 patients. We compared performance of three models: a model trained solely on the actual dataset, a fine-tuned model after pre-training, and a zero-shot learning model with only pre-training. Mean F1-score of pointwise probability plots was 0.736, 0.752, 0.761 in TD, and 0.751, 0.766, 0.772 in PD, for each model respectively. MAE was 2.19, 1.80, 1.66 in MD; 2.06, 1.90, 1.78 in PSD; and 5.63, 4.69, 4.25 in VFI. F1-score for identifying GHT normality was 0.781, 0.803, 0.785. Overall, the zero-shot model exhibited superior performance in most tasks (Figure 1).
Conclusions :
The ISP could predict most VF parameters precisely, providing valuable information on glaucoma severity. Furthermore, our findings demonstrate the potential of zero-shot learning where the globally utilized HFA could facilitate the screening ability for ISP to evaluate glaucoma.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.