Scholarlink

<p>Glaucoma is a leading cause of irreversible blindness, and early detection is critical for preventing vision loss. This paper presents HiFormer, a hybrid deep learning model that combines convolutional neural networks (CNNs) with a hierarchical attention-based transformer architecture to improve the segmentation and classification of fundus images. Unlike traditional CNN-based methods, HiFormer leverages hierarchical attention to better capture multi-scale features and subtle pathological changes associated with glaucoma. The model is trained and evaluated on both open-source datasets and clinical data from the Scientific Research Institute of Eye Diseases in Kazakhstan. Experimental results demonstrate that HiFormer outperforms standard CNN architectures, achieving a 2.91% increase in Dice coefficient and a 16.99 unit decrease in average Hausdorff Distance (HD) compared to TransUnet, and a 1.26% increase in Dice coefficient and 6.85 unit decrease in average HD compared to Swin-Unet. These metrics indicate enhanced accuracy, sensitivity, and computational efficiency, making it well-suited for real-time clinical applications. This work contributes to developing faster and more reliable diagnostic tools for glaucoma detection using retinal imaging.</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

HiFormer: a fast and accurate deep learning model for segmentation and detection of glaucoma

  • 期刊论文

  • 发表日期：

    2025年12月22日

查看全文
  • Orken Mamyrbayev,
  • Ardan Zhanegizov,
  • Kymbat Momynzhanova,
  • Alimhan Keylan

摘要

Glaucoma is a leading cause of irreversible blindness, and early detection is critical for preventing vision loss. This paper presents HiFormer, a hybrid deep learning model that combines convolutional neural networks (CNNs) with a hierarchical attention-based transformer architecture to improve the segmentation and classification of fundus images. Unlike traditional CNN-based methods, HiFormer leverages hierarchical attention to better capture multi-scale features and subtle pathological changes associated with glaucoma. The model is trained and evaluated on both open-source datasets and clinical data from the Scientific Research Institute of Eye Diseases in Kazakhstan. Experimental results demonstrate that HiFormer outperforms standard CNN architectures, achieving a 2.91% increase in Dice coefficient and a 16.99 unit decrease in average Hausdorff Distance (HD) compared to TransUnet, and a 1.26% increase in Dice coefficient and 6.85 unit decrease in average HD compared to Swin-Unet. These metrics indicate enhanced accuracy, sensitivity, and computational efficiency, making it well-suited for real-time clinical applications. This work contributes to developing faster and more reliable diagnostic tools for glaucoma detection using retinal imaging.