Multidimensional imaging model for diagnosing primary open-angle glaucoma using fundus photographs and optical coherence tomography images.

aimsTo develop a multidimensional imaging model to detect primary open-angle glaucoma (POAG) by combining fundus photographs with the retinal nerve fibre layer (RNFL) and ganglion cell-inner plexiform layer (GCIPL) thickness maps and deviation maps.

methods1054 eyes of 546 patients were included. Images were randomly divided into training, validation and test datasets at 64%, 16% and 20%. Four basic convolutional neural network algorithms were used to develop the multidimensional imaging model by combining fundus photographs, RNFL and GCIPL deviation maps, with RNFL and GCIPL thickness maps. POAG structure damage was defined as RNFL and GCIPL thinning with neuroretinal rim thinning.

resultsMultidimensional imaging model obtained an area under the receiver operating characteristic curve (AUC) of 0.970 (95% CI 0.958 to 0.979), which was superior to the fundus photographs model (AUC, 0.945 (95% CI 0.929 to 0.958); p＜0.05), RNFL and GCIPL deviation maps model (AUC, 0.931 (95% CI 0.915 to 0.946); p＜0.001), RNFL and GCIPL thickness maps model (AUC, 0.958 (95% CI 0.944 to 0.969)) in the test dataset. There was a significant difference between the multidimensional imaging model (AUC, 0.969 (95% CI 0.934 to 0.988)) and three glaucoma ophthalmologists (attending ophthalmologist 1: AUC, 0.847; attending ophthalmologist 2: AUC, 0.842; resident ophthalmologist: AUC, 0.755; p＜0.001 for both) in one fold of the test dataset. Furthermore, the overall AUCs were lower in the myopic group than in the non-myopic group. The overall AUC for lesion detection by each model (p=0.159) was: low myopia group (AUC, 0.968 (95% CI 0.940 to 0.985)) > moderate myopia group (AUC, 0.966 (95% CI 0.936 to 0.984)) >high myopia group (AUC, 0.958 (95% CI 0.927 to 0.978)). The multidimensional imaging model outperformed unimodal models in diagnosing myopia combined with POAG.

conclusionThe multidimensional imaging model, combining fundus photographs, RNFL and GCIPL deviation maps, with RNFL and GCIPL thickness maps, could perform better than unimodal models. Its potential to improve the diagnostic accuracy of myopia combined with POAG for various myopia levels is promising. This means it could be a valuable tool for assisting with POAG diagnosis in the clinic, offering hope for improved patient care.