Scholarlink

In this study, we introduce a pioneering deep learning framework, termed DR-PanopticNet, designed for the precise detection of Diabetic Retinopathy (DR), a critical ocular condition prevalent among individuals with diabetes, which poses a significant risk of vision impairment and blindness. Our framework aims to classify retinal images into five distinct categories corresponding to varying degrees of DR severity. Initially, our approach entails meticulous image pre-processing utilizing a tailored methodology integrating Gaussian blurring and enhancement techniques. Subsequently, these pre-processed images undergo analysis by various Convolutional Neural Network (CNN) architectures, serving as adept feature extractors to discern salient characteristics crucial for classification. These extracted features facilitate the categorization of DR severity across five levels, encompassing normal retinal status (absence of DR), as well as manifestations of mild, moderate, severe, and Proliferative Diabetic Retinopathy (PDR). To enhance the overall accuracy and resilience of our framework, we employ an ensemble learning strategy, denoted as DR-PanopticNet, which amalgamates the capabilities of four prominent CNN architectures: ResNet-50, DenseNet-121, InceptionV3, and EfficientNetV2. Furthermore, in addition to validating our model utilizing the established APTOS 2019 dataset, we conducted a robustness evaluation utilizing the MESSIDOR dataset. The favourable outcomes attained through this assessment underscore the efficacy and resilience of the DR-PanopticNet framework.

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

Advancing Diabetic Retinopathy Diagnosis: The Synergy of Deep Learning Models in DR-PanopticNet System

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aadi Vikram,
  • Abhinav Roy,
  • Bhavesh Gyanchandani,
  • Santosh Kumar

摘要

In this study, we introduce a pioneering deep learning framework, termed DR-PanopticNet, designed for the precise detection of Diabetic Retinopathy (DR), a critical ocular condition prevalent among individuals with diabetes, which poses a significant risk of vision impairment and blindness. Our framework aims to classify retinal images into five distinct categories corresponding to varying degrees of DR severity. Initially, our approach entails meticulous image pre-processing utilizing a tailored methodology integrating Gaussian blurring and enhancement techniques. Subsequently, these pre-processed images undergo analysis by various Convolutional Neural Network (CNN) architectures, serving as adept feature extractors to discern salient characteristics crucial for classification. These extracted features facilitate the categorization of DR severity across five levels, encompassing normal retinal status (absence of DR), as well as manifestations of mild, moderate, severe, and Proliferative Diabetic Retinopathy (PDR). To enhance the overall accuracy and resilience of our framework, we employ an ensemble learning strategy, denoted as DR-PanopticNet, which amalgamates the capabilities of four prominent CNN architectures: ResNet-50, DenseNet-121, InceptionV3, and EfficientNetV2. Furthermore, in addition to validating our model utilizing the established APTOS 2019 dataset, we conducted a robustness evaluation utilizing the MESSIDOR dataset. The favourable outcomes attained through this assessment underscore the efficacy and resilience of the DR-PanopticNet framework.