Scholarlink

Diabetic retinopathy (DR) is a prevalent impediment of diabetes marked by irregular blood vessel development in the eye’s retina, potentially resulting in blindness if not properly managed. Conventional diabetic retinopathy screening by eye doctors takes up a lot of time, underscoring the necessity for improved detection techniques. This research investigates employing deep learning (DL) for automating DR detection by utilizing MobileNet, a compact model trained on 3,662 high-resolution fundus images from the Diabetic retinopathy Gaussian filtered dataset. The MobileNet model categorizes DR into five different levels, from 0 to 4. In order to enhance precision, we explore hybrid models that combine MobileNet with different architectures. The MobileNet + Recurrent Neural Network (RNN) combination aims to highlight the connections between sequences and time intervals, improving the model’s capacity to identify subtle changes in DR stages. On the contrary, the fusion of MobileNet with Graph Neural Network (GNN) recognizes detailed spatial designs in retinal pictures, enhancing the model’s capability to analyze intricate structural characteristics. Our findings indicate that the MobileNet + GNN model had a higher accuracy when compared to the MobileNet + RNN model. These results show that combining spatial and temporal feature extraction in hybrid models can greatly enhance DR detection, providing a more effective option compared to conventional methods. This method shows a hopeful advancement in automatic, immediate DR screening within clinical environments.

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

A Lightweight Deep Learning Approach for Diabetic Retinopathy Detection: MobileNet with RNN and GNN Hybrid Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • K. Lakshmi Prasanna,
  • G. Kumari

摘要

Diabetic retinopathy (DR) is a prevalent impediment of diabetes marked by irregular blood vessel development in the eye’s retina, potentially resulting in blindness if not properly managed. Conventional diabetic retinopathy screening by eye doctors takes up a lot of time, underscoring the necessity for improved detection techniques. This research investigates employing deep learning (DL) for automating DR detection by utilizing MobileNet, a compact model trained on 3,662 high-resolution fundus images from the Diabetic retinopathy Gaussian filtered dataset. The MobileNet model categorizes DR into five different levels, from 0 to 4. In order to enhance precision, we explore hybrid models that combine MobileNet with different architectures. The MobileNet + Recurrent Neural Network (RNN) combination aims to highlight the connections between sequences and time intervals, improving the model’s capacity to identify subtle changes in DR stages. On the contrary, the fusion of MobileNet with Graph Neural Network (GNN) recognizes detailed spatial designs in retinal pictures, enhancing the model’s capability to analyze intricate structural characteristics. Our findings indicate that the MobileNet + GNN model had a higher accuracy when compared to the MobileNet + RNN model. These results show that combining spatial and temporal feature extraction in hybrid models can greatly enhance DR detection, providing a more effective option compared to conventional methods. This method shows a hopeful advancement in automatic, immediate DR screening within clinical environments.