Scholarlink

Background <p>Diabetic retinopathy (DR) is the primary cause of visual impairments impacting the diabetic population worldwide. Early detection and classification of diabetic retinopathy (DR) are crucial for preventing vision impairment and blindness in diabetic patients.</p> Objective <p>With the aim of improving the fine-grained DR classification, a similarity learning model, based on robust lesion-aware features and few-shot learning, is proposed in this work.</p> Method <p>A lightweight autoencoder is designed to enhance the detection of small lesions. It integrates channel and spatial attention layers, enabling the network to efficiently focus on lesion-relevant regions. N-shot meta-learning is applied for severity classification under limited data conditions. The proposed model is trained and assessed on six publicly available fundus image datasets. To evaluate adaptability to regional variations, the model is validated on a curated Indian fundus dataset comprising 100 well-labeled fundus images.</p> Results <p>NetraNet is evaluated on all the key metrics that are necessary for medical diagnosis, obtaining the detection accuracy of 99.5%, 0.993 precision, and 0.994 recall. This high level of performance demonstrates the potential of NetraNet in DR diagnosis. For multi-class classification, the model achieves an accuracy of 95.9% with a Cohen’s kappa score of 0.95, indicating high performance and robustness. Our model attains a significant Kappa score of 0.94 on collected data, serving as a reliable tool for clinical DR screening.</p> Conclusion <p>NetraNet exhibits strong performance across diverse datasets, even in limited data conditions, making it a favorable method for real-world clinical screening scenarios where well-labeled data is often limited. This work demonstrates the impact of strong feature extraction along with similarity learning models in medical image classification and offers a scalable solution for other fundus disease detection.</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

NetraNet: A similarity learning method for diabetic retinopathy classification and clinical diagnosis

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Preeti Kapoor,
  • Shaveta Arora

摘要

Background

Diabetic retinopathy (DR) is the primary cause of visual impairments impacting the diabetic population worldwide. Early detection and classification of diabetic retinopathy (DR) are crucial for preventing vision impairment and blindness in diabetic patients.

Objective

With the aim of improving the fine-grained DR classification, a similarity learning model, based on robust lesion-aware features and few-shot learning, is proposed in this work.

Method

A lightweight autoencoder is designed to enhance the detection of small lesions. It integrates channel and spatial attention layers, enabling the network to efficiently focus on lesion-relevant regions. N-shot meta-learning is applied for severity classification under limited data conditions. The proposed model is trained and assessed on six publicly available fundus image datasets. To evaluate adaptability to regional variations, the model is validated on a curated Indian fundus dataset comprising 100 well-labeled fundus images.

Results

NetraNet is evaluated on all the key metrics that are necessary for medical diagnosis, obtaining the detection accuracy of 99.5%, 0.993 precision, and 0.994 recall. This high level of performance demonstrates the potential of NetraNet in DR diagnosis. For multi-class classification, the model achieves an accuracy of 95.9% with a Cohen’s kappa score of 0.95, indicating high performance and robustness. Our model attains a significant Kappa score of 0.94 on collected data, serving as a reliable tool for clinical DR screening.

Conclusion

NetraNet exhibits strong performance across diverse datasets, even in limited data conditions, making it a favorable method for real-world clinical screening scenarios where well-labeled data is often limited. This work demonstrates the impact of strong feature extraction along with similarity learning models in medical image classification and offers a scalable solution for other fundus disease detection.