Scholarlink

&#xa0;Background <p>Diabetic retinopathy (DR) is an ocular condition affecting the retina’s blood vessels. Patients with diabetes who experience DR risk losing all vision. In this situation, DR must be detected early to save the vision and provide prompt treatment. DR manual diagnosis is time-consuming and prone to mistakes due to its complexities. Numerous machine/deep learning techniques have been suggested to identify DR and its various stages from retinal images automatically.</p> Objective <p>This study presents the deep learning model for automatically classifying DR stages from fundus images.&#xa0;</p> Methods <p>This work proposes a transformer and parallel ensemble-based deep learning model for diabetic retinopathy segmentation and detection called TPEDLM. The proposed work contains three steps: preprocessing, segmentation, and detection. The fundus images are enhanced using contrast-limited adaptive histogram equalization in the preprocessing step. The TransUNet model segments the fundus image that extracts retina blood vessels, optic disc, and DR lesions. The three deep learning models, Inception V3, ResNet50 and VGG16, are used to detect the DR stages.</p> Results <p>The IDRiD and Messidor-2 datasets are used to evaluate the proposed methodology. The experimental results confirm that the proposed method effectively distinguishes the different stages of DR, achieving over 98% accuracy across both datasets. However, external validation on prospectively collected, multi-centre datasets is necessary before clinical deployment.</p> Conclusions <p>The suggested methodology has exhibited enhanced efficacy across many parameters for the classification of diabetic retinopathy using retinal fundus pictures. Thus, the model can function as a crucial tool for professionals in the early identification and management of diabetic patients.</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

Transformer and parallel ensemble-based deep learning model for Diabetic Retinopathy segmentation and detection

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Syedzagiriya S,
  • MD ZUBAIR RAHMAN A M J

摘要

 Background

Diabetic retinopathy (DR) is an ocular condition affecting the retina’s blood vessels. Patients with diabetes who experience DR risk losing all vision. In this situation, DR must be detected early to save the vision and provide prompt treatment. DR manual diagnosis is time-consuming and prone to mistakes due to its complexities. Numerous machine/deep learning techniques have been suggested to identify DR and its various stages from retinal images automatically.

Objective

This study presents the deep learning model for automatically classifying DR stages from fundus images. 

Methods

This work proposes a transformer and parallel ensemble-based deep learning model for diabetic retinopathy segmentation and detection called TPEDLM. The proposed work contains three steps: preprocessing, segmentation, and detection. The fundus images are enhanced using contrast-limited adaptive histogram equalization in the preprocessing step. The TransUNet model segments the fundus image that extracts retina blood vessels, optic disc, and DR lesions. The three deep learning models, Inception V3, ResNet50 and VGG16, are used to detect the DR stages.

Results

The IDRiD and Messidor-2 datasets are used to evaluate the proposed methodology. The experimental results confirm that the proposed method effectively distinguishes the different stages of DR, achieving over 98% accuracy across both datasets. However, external validation on prospectively collected, multi-centre datasets is necessary before clinical deployment.

Conclusions

The suggested methodology has exhibited enhanced efficacy across many parameters for the classification of diabetic retinopathy using retinal fundus pictures. Thus, the model can function as a crucial tool for professionals in the early identification and management of diabetic patients.