Application performance and limitations of general and ophthalmology-specific large models in ophthalmic images
10.3980/j.issn.1672-5123.2026.9.24
- VernacularTitle:通用和眼科专用大模型在眼科图像中的应用效果及局限
- Author:
Shixin LAI
1
;
Xinyan FAN
1
;
Mingjie LUO
1
Author Information
1. State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-sen University;Guangdong Provincial Key Laboratory of Ophthalmology and Visual Science, Guangzhou 510623, Guangdong Province, China
- Publication Type:Journal Article
- Keywords:
vision-language models;
artificial intelligence;
ophthalmology;
application performance;
evaluation
- From:
International Eye Science
2026;26(9):1651-1657
- CountryChina
- Language:Chinese
-
Abstract:
With the rapid advancement of large language models, both general-purpose and ophthalmology-specific variants have shown substantial potential for ophthalmic image analysis. However, their task-specific effectiveness and applicable boundaries in clinical practice remain to be fully clarified. This review outlines the technical evolution of large models in ophthalmology, including masked visual modeling, vision-language contrastive learning, and subsequent stages of knowledge integration and multimodal reasoning. The performance of general-purpose versus ophthalmology-specific models across common clinical scenarios, including disease screening and initial triage, disease grading and lesion segmentation, and medical report generation with multimodal reasoning were further compared. Results show that general-purpose models are well-suited for multi-disease screening and primary-care triage, whereas ophthalmology-specific models excel in fine-grained grading, subtle lesions detection, and rare disease diagnosis. Moreover, instruction-tuned domain-specific models produce medical reports with greater clinical consistency. These findings provide practical guidance for selecting appropriate model types and technical pathways tailored to specific ophthalmic tasks. Furthermore, the review identifies the current challenges facing these models, including reliability, generalizability, evaluation frameworks, and reproducibility, as well as the resulting limitations in clinical application.
