Scholarlink

<p>Deep learning has significantly advanced medical imaging, especially in ophthalmology, by facilitating automated diagnostic tools to address critical clinical challenges. This research introduces a deep learning based approach for the semantic segmentation of diabetic retinopathy (DR) lesions, which is vital to early identification and intervention. Conventional segmentation techniques mostly utilize Convolutional Neural Networks (CNNs), Fully Convolutional Networks (FCNs), and UNET architectures, frequently using transfer learning to address dataset constraints. This study substitutes the traditional UNET encoder with MobileNetV2, a lightweight and efficient option for feature extraction and lesion segmentation. Experimental results indicate enhanced efficacy in the segmentation of hemorrhages and soft exudates, with sensitivity scores of 0.89 and 0.97, respectively. This study introduces lesion-level performance analysis, which offers a more clinically meaningful assessment than conventional pixel-based segmentation. Unlike pixel-wise methods, lesion-based evaluation aligns with how ophthalmologists diagnose DR, making AI-based screening tools more interpretable and reliable. The suggested model equilibrates accuracy, efficiency, and computational viability, rendering it appropriate for real-time clinical applications and mobile-based diabetic retinopathy screening tools.</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

Semantic Segmentation of Diabetic Retinopathy Lesions Using Deep Learning

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Dimitrios Theodoropoulos,
  • Nikolaos Sifakis,
  • Georgios Manikis,
  • Giorgos Papadourakis,
  • Konstantinos Armyras,
  • Konstantinos Marias

摘要

Deep learning has significantly advanced medical imaging, especially in ophthalmology, by facilitating automated diagnostic tools to address critical clinical challenges. This research introduces a deep learning based approach for the semantic segmentation of diabetic retinopathy (DR) lesions, which is vital to early identification and intervention. Conventional segmentation techniques mostly utilize Convolutional Neural Networks (CNNs), Fully Convolutional Networks (FCNs), and UNET architectures, frequently using transfer learning to address dataset constraints. This study substitutes the traditional UNET encoder with MobileNetV2, a lightweight and efficient option for feature extraction and lesion segmentation. Experimental results indicate enhanced efficacy in the segmentation of hemorrhages and soft exudates, with sensitivity scores of 0.89 and 0.97, respectively. This study introduces lesion-level performance analysis, which offers a more clinically meaningful assessment than conventional pixel-based segmentation. Unlike pixel-wise methods, lesion-based evaluation aligns with how ophthalmologists diagnose DR, making AI-based screening tools more interpretable and reliable. The suggested model equilibrates accuracy, efficiency, and computational viability, rendering it appropriate for real-time clinical applications and mobile-based diabetic retinopathy screening tools.