Scholarlink

<p>Artificial intelligence (AI) has emerged as a key component of modern ophthalmology, enabling highly precise diagnosis of various posterior segment pathologies. Yet, the development of AI technology is inconsistent between subspecialties. Research on AI technology is dominated by studies on retinal diseases, such as diabetic retinopathy (DR) and age-related macular degeneration (AMD), enabled by DL models and datasets like IDRiD. Classification (46.5%), segmentation (41.9%), and detection (4.7%) models are being used more and more in DR studies. YOLO, Vision Transformers, and U-Net derivatives are the most common types. Significant constraints still exist despite quantifiable performance improvements, such as limited external generalizability, class imbalance, and a lack of high-quality labeled data. Convolutional neural networks have been shown to have strong diagnostic and predictive capabilities in macular hole (MH) research, but therapeutic decision support is still largely unexplored. Multimodal fusion of fundus imaging, OCT, and clinical parameters improves glaucoma studies, but explainability is applied inconsistently. On the other hand, because of inconsistent data standards and small datasets, anterior segment and neuro-ophthalmic disorders continue to be underrepresented. Although bedside validation is still limited, emerging eXplainable AI (XAI) and generative AI offer new ways to address transparency and data scarcity. Generative models are increasingly used for synthetic augmentation and early decision-support tasks. A promising route toward clinically interpreted, reliable, and deployable ophthalmic AI systems is the integration of XAI and generative frameworks with multimodal architectures.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable and Generative AI in Ophthalmology: Progress, Gaps, and Future Directions Beyond the Retina

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Parijata Majumdar,
  • Sanjoy Mitra

摘要

Artificial intelligence (AI) has emerged as a key component of modern ophthalmology, enabling highly precise diagnosis of various posterior segment pathologies. Yet, the development of AI technology is inconsistent between subspecialties. Research on AI technology is dominated by studies on retinal diseases, such as diabetic retinopathy (DR) and age-related macular degeneration (AMD), enabled by DL models and datasets like IDRiD. Classification (46.5%), segmentation (41.9%), and detection (4.7%) models are being used more and more in DR studies. YOLO, Vision Transformers, and U-Net derivatives are the most common types. Significant constraints still exist despite quantifiable performance improvements, such as limited external generalizability, class imbalance, and a lack of high-quality labeled data. Convolutional neural networks have been shown to have strong diagnostic and predictive capabilities in macular hole (MH) research, but therapeutic decision support is still largely unexplored. Multimodal fusion of fundus imaging, OCT, and clinical parameters improves glaucoma studies, but explainability is applied inconsistently. On the other hand, because of inconsistent data standards and small datasets, anterior segment and neuro-ophthalmic disorders continue to be underrepresented. Although bedside validation is still limited, emerging eXplainable AI (XAI) and generative AI offer new ways to address transparency and data scarcity. Generative models are increasingly used for synthetic augmentation and early decision-support tasks. A promising route toward clinically interpreted, reliable, and deployable ophthalmic AI systems is the integration of XAI and generative frameworks with multimodal architectures.