Table 1 Recent applications of AI in ophthalmology

Szanto et al., 2025

[3].

Optic nerve head (ONH) atrophy classification: glaucoma, NAION, optic neuritis, and healthy controls

Three ResNet‑3D‑18 models: whole volume, peripapillary region only, ONH only; 5‑fold stratified cross‑validation; unsegmented 3D OCT volumes

Multicenter; Device: Cirrus ONH OCT; 7,014 scans/1,382 eyes. Classes: glaucoma n = 113, NAION n = 311, optic neuritis n = 163, healthy n = 715. Mean age 54.2 ± 16.9; 65% male, 35% female

Whole‑volume: Accuracy 88.9%; Macro AUC‑ROC 0.977 (95% CI 0.974–0.979); F1 (glaucoma 0.94, NAION 0.87, optic neuritis 0.78, healthy 0.91). Peripapillary region:

Accuracy 85.9%; AUC 0.970 (0.967–0.972). ONH:

Accuracy 87.0%; AUC 0.972 (0.970–0.975).

Madadi et al., 2025 [4].

Neuro‑ophthalmic diseases (varied chronic and acute cases; diagnosis from case reports)

LLM approach: ChatGPT (GPT‑3.5 and GPT‑4); each case entered as a prompt to obtain the most likely diagnosis

22 neuro‑ophthalmic case reports from a public database; same information presented to 2 neuro‑ophthalmologists (NO); responses compared

Accuracy: GPT‑3.5 = 13/22 (59%), GPT‑4 = 18/22 (82%), NOs = 19/22 (86%). Agreement: GPT‑4 & NO‑1 = 17/22 (77%); GPT‑4 & NO‑2 = 16/22 (73%); GPT‑3.5 & GPT‑4 = 13/22 (59%); NO‑1 & NO‑2 = 17/22 (77%).

Gu et al., 2025

[5].

Neuromyelitis optica (NMO): onset and stage prediction

Multimodal AI integrating ultrawide‑field fundus (UWF) photographs with clinical examination data

UWF of 330 eyes from 285 NMO patients + 1,288 eyes from 770 non‑NMO participants; clinical exam reports included

AUC: 0.9923; Max Youden index: 0.9389; Sensitivity: 97.0%; Specificity: 96.9% (test dataset for NMO prevalence prediction).

Gungor et al., 2025

[6].

Papilledema and other optic neuropathies (optic disc abnormalities)

CNN‑based DL System: optic disc and peripapillary region segmentation; 3‑class classification (papilledema/other abnormalities/normal)

International, multiethnic, multicenter: 20,533 retinal photos (10,647 patients) from 31 centers. Trai + internal validation: 18,981 mydriatic photos (9,830 patients; 22 countries). External test: 1,552 prospective photos (817 patients) with a handheld nonmydriatic camera (Aurora, Optomed, Finland). Outcomes evaluated at eye and patient levels.

For differentiating papilledema from others and healthy: Accuracy 99.5% (95% CI 99.1–99.8), Sensitivity 81.0% (95% CI 64.1–97.7), Specificity 99.7% (95% CI 99.5–99.9), AUC 98.3% (95% CI 97.6–99.9).

Szanto et al., 2025

[7].

Papilledema vs. NAION vs. healthy

DL (ResNet‑50 backbone); disc localization normalization; Grad‑CAM explainability; 5‑fold cross‑validation with majority voting for external validation

Train/validation: 15,088 fundus photos from 5,866 eyes; External validation: 1,126 images from 928 eyes

Internal validation: Accuracy ≈ 96.2%, Macro AUC ≈ 0.995; External validation: Accuracy ≈ 93.6%, Macro AUC ≈ 0.980