Scholarlink

<p>Diabetic Retinopathy (DR) is a principal cause of preventable blindness in the developed world. The increasing worldwide pandemic of diabetes and a shortage of ophthalmologists have heightened the need for automatic, robust early-stage screeners. Over the past decade, research has shifted away from traditional machine-learning (hand-crafted) pipelines to deep or transformer-based models that can learn intricate retinal representations directly from fundus images. This PRISMA-compliant systematic review analyses 50 papers through IEEE Xplore, PubMed, Scopus, WoS (Web of Science), ScienceDirect and Springer from 2016 to 2026. In addition to quantitative analysis of ML, CNN, and transformer-based models, including accuracy, sensitivity, specificity, and AUC, a bibliometric examination was performed to study publication trends, research collaborations, and citation trends. The findings indicate a shift from conventional classifiers (SVM, Random Forest) to end-to-end CNN-based models and transposed models (e.g., Swin Transformer, EfficientNetV2, Vision Transformer). The contributions were mainly concentrated in India, China, and the USA, and there was an increasing demand for an interdisciplinary interface between engineering disciplines with biomedical sciences. While state-of-the-art algorithms achieve 94–95% accuracy, their generalizability is limited to 85–93% on a generalized test dataset, and issues such as data imbalance, lack of interpretability, and clinical validation remain. With many advanced AI systems across the board of innovation, explainable, resource-efficient, and multimodal platforms integrating fundus photographic data with patient-centered clinical information should set their sights on equitable 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

Artificial Intelligence in Diabetic Retinopathy Detection: A Decade of Progress from Machine Learning to Transformer-Based Frameworks

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Vijaya Kumar Sagenela,
  • Karthika Gidijala,
  • Santoshachandra Rao Karanam,
  • N. Shyam Sunder Sagar

摘要

Diabetic Retinopathy (DR) is a principal cause of preventable blindness in the developed world. The increasing worldwide pandemic of diabetes and a shortage of ophthalmologists have heightened the need for automatic, robust early-stage screeners. Over the past decade, research has shifted away from traditional machine-learning (hand-crafted) pipelines to deep or transformer-based models that can learn intricate retinal representations directly from fundus images. This PRISMA-compliant systematic review analyses 50 papers through IEEE Xplore, PubMed, Scopus, WoS (Web of Science), ScienceDirect and Springer from 2016 to 2026. In addition to quantitative analysis of ML, CNN, and transformer-based models, including accuracy, sensitivity, specificity, and AUC, a bibliometric examination was performed to study publication trends, research collaborations, and citation trends. The findings indicate a shift from conventional classifiers (SVM, Random Forest) to end-to-end CNN-based models and transposed models (e.g., Swin Transformer, EfficientNetV2, Vision Transformer). The contributions were mainly concentrated in India, China, and the USA, and there was an increasing demand for an interdisciplinary interface between engineering disciplines with biomedical sciences. While state-of-the-art algorithms achieve 94–95% accuracy, their generalizability is limited to 85–93% on a generalized test dataset, and issues such as data imbalance, lack of interpretability, and clinical validation remain. With many advanced AI systems across the board of innovation, explainable, resource-efficient, and multimodal platforms integrating fundus photographic data with patient-centered clinical information should set their sights on equitable DR screening.