Deep learning-based prediction of 24-2 visual field from fundus photographs in glaucoma

To develop a deep learning architecture for predicting visual field (VF) maps from fundus images. Fundus images and VF test results were obtained from two tertiary hospitals. The training dataset comprised 31,443 fundus photographs and 27,364 VF examinations, whereas the test dataset comprised 6,436 fundus photographs and 5,647 VF examinations. A novel deep learning architecture, TDV-Net, was developed to predict VF maps and assess glaucoma progression from input fundus images. The model performance was evaluated using the pointwise mean absolute error (MAE) between the predicted and actual VF. Additionally, the ability of the model to predict glaucoma progression was compared with that of human experts. The overall prediction error (MAE) of TDV-Net was 3.91 dB across all glaucoma stages. According to glaucoma severity, the MAEs were 3.09, 5.66, and 9.15 dB for mild, moderate, and severe glaucoma, respectively. In predicting glaucoma progression, TDV-Net achieved an accuracy of 76.3%, sensitivity of 73.4%, and specificity of 77.1%, outperforming human experts. TDV-Net demonstrated robust performance in predicting the VFs from fundus photographs and outperformed human experts in predicting glaucoma progression. TDV-Net has the potential to enhance accessibility to VF assessment, particularly in settings where standard perimetric testing is not readily available, such as routine health checkups or primary care environments.