Scholarlink

<p>Diabetic retinopathy is a significant cause of vision impairment and blindness, making early detection and accurate classification essential to mitigate its progression and enable timely treatment. With the global prevalence of diabetes increasing dramatically, there is a growing demand for automated, precise, and efficient diabetic retinopathy (DR) detection systems. Developing robust detection frameworks remains challenging due to the complex variations in retinal images. In addition, fine-grained classification across different DR stages—standard, mild, moderate, severe, and proliferative—remains a critical requirement. This paper proposes a novel hybrid framework that integrates Unified Q-learning-based GoogleNet (UniQ-GNet) with slime-whale optimization (SliWha Opt) for DR detection and classification. Inspired by the intelligent foraging behavior of slime moulds and cooperative hunting patterns of whales, SliWha Opt optimizes the architectural depth of GoogleNet, enhancing feature representation while simultaneously reducing model complexity and computational overhead—key limitations of existing deep learning models. The originality of this work lies in the seamless integration of reinforcement learning with bio-inspired optimization, enabling the model to be adaptive and self-regulating, thereby improving DR classification accuracy. Extensive experimentation on the Kaggle EyePACS and APTOS 2019 blindness detection datasets demonstrated the effectiveness of the proposed framework. The model achieved state-of-the-art performance, with accuracy scores of 98.85% on the Kaggle dataset and 98.79% on the APTOS dataset, outperforming benchmark models such as DenseNet, VGG-19, and Inception networks. A detailed performance evaluation across different DR severity levels consistently confirmed high precision, recall, and F1-score values exceeding 98%. In addition, the proposed approach demonstrated superior computational efficiency, processing Kaggle images in 12&#xa0;s and APTOS images in 14&#xa0;s, significantly outperforming traditional machine learning techniques such as SVM, KNN, and Random Forests. This study addresses key challenges in medical image analysis and highlights the potential for real-time clinical applications. Moreover, it paves the way for future research to extend this approach to other ophthalmological diseases and advance healthcare automation. However, challenges related to model interpretability, real-world applicability, and potential biases must be addressed for the model to be widely adopted in clinical settings. Future work should focus on integrating explainability techniques, validating the model on real-world clinical data, and adapting it for diverse healthcare settings to enhance its practical usability and trustworthiness.</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

Reinforced vision model using Q-learning GoogleNet with slime-whale optimization for diabetic retinopathy detection

  • 期刊论文

  • 发表日期：

    2025年7月12日

查看全文
  • Vijayalaxmi Gopu,
  • M. Selvi

摘要

Diabetic retinopathy is a significant cause of vision impairment and blindness, making early detection and accurate classification essential to mitigate its progression and enable timely treatment. With the global prevalence of diabetes increasing dramatically, there is a growing demand for automated, precise, and efficient diabetic retinopathy (DR) detection systems. Developing robust detection frameworks remains challenging due to the complex variations in retinal images. In addition, fine-grained classification across different DR stages—standard, mild, moderate, severe, and proliferative—remains a critical requirement. This paper proposes a novel hybrid framework that integrates Unified Q-learning-based GoogleNet (UniQ-GNet) with slime-whale optimization (SliWha Opt) for DR detection and classification. Inspired by the intelligent foraging behavior of slime moulds and cooperative hunting patterns of whales, SliWha Opt optimizes the architectural depth of GoogleNet, enhancing feature representation while simultaneously reducing model complexity and computational overhead—key limitations of existing deep learning models. The originality of this work lies in the seamless integration of reinforcement learning with bio-inspired optimization, enabling the model to be adaptive and self-regulating, thereby improving DR classification accuracy. Extensive experimentation on the Kaggle EyePACS and APTOS 2019 blindness detection datasets demonstrated the effectiveness of the proposed framework. The model achieved state-of-the-art performance, with accuracy scores of 98.85% on the Kaggle dataset and 98.79% on the APTOS dataset, outperforming benchmark models such as DenseNet, VGG-19, and Inception networks. A detailed performance evaluation across different DR severity levels consistently confirmed high precision, recall, and F1-score values exceeding 98%. In addition, the proposed approach demonstrated superior computational efficiency, processing Kaggle images in 12 s and APTOS images in 14 s, significantly outperforming traditional machine learning techniques such as SVM, KNN, and Random Forests. This study addresses key challenges in medical image analysis and highlights the potential for real-time clinical applications. Moreover, it paves the way for future research to extend this approach to other ophthalmological diseases and advance healthcare automation. However, challenges related to model interpretability, real-world applicability, and potential biases must be addressed for the model to be widely adopted in clinical settings. Future work should focus on integrating explainability techniques, validating the model on real-world clinical data, and adapting it for diverse healthcare settings to enhance its practical usability and trustworthiness.