Scholarlink

<p>Significant progress has been made in the automatic recognition of eye diseases. However, existing eye disease detection models face several challenges, including dataset imbalance, difficulty in identifying early-stage lesions, insufficient model generalization, and the need for clinical real-time performance. To address these issues, this study collaborated with an ophthalmology clinic to construct a dataset containing 4,204 retinal images, thereby filling the gap in the available datasets for eye disease detection. Based on this dataset, a novel eye disease detection model, EYE-YOLOv10s (You Only Look Once), is proposed. This model, built upon the YOLOv10s framework, introduces the NAM (Normalization-based Attention Module) and SOCA (Second-order Channel Attention) attention mechanisms. SOCA enhances the covariance relationship between different channels, improving the ability to detect subtle lesions, while NAM assigns different weights to channels and spatial features, emphasizing critical features and suppressing background noise. Additionally, the Ghost module and C2f_Faster (CSP Bottleneck with 2 Convolutions) module are incorporated to optimize the backbone network, improving detection performance and reducing computational complexity. Finally, an improved loss function is employed, incorporating a penalty term based on the center point distance of the bounding boxes, which increases localization accuracy and robustness in low-quality images. Experimental results demonstrate that EYE-YOLOv10s achieves mAP@50 (mean Average Precision at 50% Intersection over Union) values of 94.7% and 98.5% on two independent eye datasets, surpassing YOLOv10s by 2.83% and 3.14%, respectively. At the same time, the model reduces the number of parameters and floating-point operations by 8.7% and 5.31%, respectively. <a href="https://github.com/fe-52/EYE-yolov10s/tree/master">https://github.com/fe-52/EYE-yolov10s/tree/master</a>.</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

Research on the application of the improved YOLOv10s algorithm in eye disease detection

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Hongfang Ma,
  • Lihong Chang,
  • Renhao Jiao,
  • Jin Dang,
  • Xiaoli Luo,
  • Xue Bai

摘要

Significant progress has been made in the automatic recognition of eye diseases. However, existing eye disease detection models face several challenges, including dataset imbalance, difficulty in identifying early-stage lesions, insufficient model generalization, and the need for clinical real-time performance. To address these issues, this study collaborated with an ophthalmology clinic to construct a dataset containing 4,204 retinal images, thereby filling the gap in the available datasets for eye disease detection. Based on this dataset, a novel eye disease detection model, EYE-YOLOv10s (You Only Look Once), is proposed. This model, built upon the YOLOv10s framework, introduces the NAM (Normalization-based Attention Module) and SOCA (Second-order Channel Attention) attention mechanisms. SOCA enhances the covariance relationship between different channels, improving the ability to detect subtle lesions, while NAM assigns different weights to channels and spatial features, emphasizing critical features and suppressing background noise. Additionally, the Ghost module and C2f_Faster (CSP Bottleneck with 2 Convolutions) module are incorporated to optimize the backbone network, improving detection performance and reducing computational complexity. Finally, an improved loss function is employed, incorporating a penalty term based on the center point distance of the bounding boxes, which increases localization accuracy and robustness in low-quality images. Experimental results demonstrate that EYE-YOLOv10s achieves mAP@50 (mean Average Precision at 50% Intersection over Union) values of 94.7% and 98.5% on two independent eye datasets, surpassing YOLOv10s by 2.83% and 3.14%, respectively. At the same time, the model reduces the number of parameters and floating-point operations by 8.7% and 5.31%, respectively. https://github.com/fe-52/EYE-yolov10s/tree/master.