Abstract
Purpose :
To assess the potential of deep learning techniques to predict visual field (VF) progression. We trained and validated a convolutional neural network (CNN) with a baseline fundus photo as input, to predict Mean Deviation (MD) rate of progression.
Methods :
Data were retrospectively collected from the glaucoma clinic at the University Hospitals UZ Leuven, Belgium. Inclusion criteria were defined as having at least five reliable VF exams of the same type, collected over a period of at least 2.5 years. Reliable VF exams were defined as either a fixation loss ≤20% for Humphrey Field Analyzer (HFA) 24-2 exams, or a false positive rate ≤15% for Octopus G1 exams (only in training set). Eligible fundus photos had to be captured within 200 days of the initial VF exam. Progression was defined as regressed MD rate (dB/year), with a cutoff between -0.5 and -1.5, coupled with a significant negative slope (p<0.1). Regression was obtained using ordinary least squares (OLS) or Huber robust regression (Fig 1). The eligible set consisted of 2382 eyes from 1527 individuals that were followed up and treated for glaucoma.
Preprocessed fundus images were used as input to a custom ResNet-50 CNN that was adapted to the prediction of VF progression. In total, 12 experiments were conducted: six using OLS, and six using Huber. Each experiment ran for 10 epochs (77 iterations with a batch size of 24), with the model weights averaged over the last five epochs. Model performance was evaluated using the area under the receiver operating characteristic curve [AUC, (95% CI)].
Results :
On average, eyes were followed up for more than five years, and had more than eight VF exams (Fig 2, Study Sample Info). The prevalence of progressors ranged from 4% to 15%, depending on the dB/y cutoff and regression type.
Using OLS regression and a cutoff at -0.5 yielded the best AUC [0.67 (0.58-0.76)] on the validation set of eyes with HFA exams (Fig 2, Results). Remarkably, only two experiments resulted in a significant AUC on the test set. The model with Huber regression, with a cutoff at -0.7, obtained an AUC of 0.65 (0.55-0.70).
Conclusions :
Deep learning models can predict VF progressors from a baseline fundus photo with moderate accuracy. The potential effect of glaucoma treatment on this study outcome warrants further investigation. Future experiments should assess the added value of including fundus photos from multiple visits.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.