A multi-label transformer-based deep learning approach to predict focal visual field progression.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Standard automated perimetry is the current clinical standard for tracking visual field (VF) changes in glaucoma diagnosis.
• Predicting regional VF progression using advanced methods remains an underexplored area in glaucoma research.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model for predicting regional visual field (VF) progression in glaucoma patients.
• To assess the model's ability to identify and forecast progression in specific VF regions using longitudinal data.

Main Methods:

• A multi-label transformer-based network (MTN) was developed using longitudinal VF data from 2430 eyes of 1283 glaucoma patients.
• Progression was defined by the mean deviation (MD) slope across six VF regions (clusters).
• MTN models were trained and tested for focal progression detection and forecasting using varying numbers of VFs as input.

Main Results:

• MTNs demonstrated excellent performance (macro-average AUC > 0.884) in detecting focal VF progression with five or more VFs.
• Using six VFs improved model performance and stability compared to five VFs, achieving a macro-average AUC of 0.848 for forecasting.
• The model showed excellent performance (AUC ≥ 0.86) in four clusters for eyes with severe VF loss (baseline MD ≤ -12 dB).

Conclusions:

• The developed multi-label DL networks show high prediction accuracy for identifying and forecasting VF progression.
• These DL models hold potential as assistive tools for clinicians in managing glaucoma progression.