Scholarlink

<p>Diabetic retinopathy (DR) remains a leading cause of preventable blindness worldwide, placing a growing burden on healthcare systems. While artificial intelligence (AI) offers promising tools for early detection and scalable screening, its transition from research to real-world clinical use faces significant hurdles. This systematic review, guided by PRISMA principles, examines 27 recent studies (2023–2025) to map the evolving landscape of AI-driven DR diagnosis. We categorize approaches into five families: convolutional neural networks (CNNs), hybrid CNN–machine learning models, transformer-based architectures, clinical data–driven predictors, and multimodal fusion systems. Our analysis reveals that while transformers excel in severity grading and hybrid models demonstrate practical robustness, critical gaps persist—including poor early-DR sensitivity, limited generalizability across imaging devices, and a lack of clinical explainability. We argue that future efforts must prioritize lightweight, interpretable, and temporally aware AI systems that integrate multimodal patient data. By bridging technical innovation with clinical pragmatism, this review aims to accelerate the development of deployable AI solutions for global DR screening.</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

Hybrid and transformer based artificial intelligence for diabetic retinopathy diagnosis a systematic review of methods challenges and clinical readiness

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Frenisha Digaswala,
  • Amit Ganatra

摘要

Diabetic retinopathy (DR) remains a leading cause of preventable blindness worldwide, placing a growing burden on healthcare systems. While artificial intelligence (AI) offers promising tools for early detection and scalable screening, its transition from research to real-world clinical use faces significant hurdles. This systematic review, guided by PRISMA principles, examines 27 recent studies (2023–2025) to map the evolving landscape of AI-driven DR diagnosis. We categorize approaches into five families: convolutional neural networks (CNNs), hybrid CNN–machine learning models, transformer-based architectures, clinical data–driven predictors, and multimodal fusion systems. Our analysis reveals that while transformers excel in severity grading and hybrid models demonstrate practical robustness, critical gaps persist—including poor early-DR sensitivity, limited generalizability across imaging devices, and a lack of clinical explainability. We argue that future efforts must prioritize lightweight, interpretable, and temporally aware AI systems that integrate multimodal patient data. By bridging technical innovation with clinical pragmatism, this review aims to accelerate the development of deployable AI solutions for global DR screening.