Scholarlink

<p>Accurate segmentation of retinal blood vessels in fundus images is beneficial for early diagnosis and prevention of retinal diabetic and cardiovascular diseases. However, accurate segmentation of retinal blood vessels is a challenging task due to issues such as small vessels, vascular lesions and uneven background brightness. With the rapid development of computer vision technology and the rise of deep neural networks, the U-Net has demonstrated good performance in solving retinal vessel segmentation problems. However, the classic U-Net structure is limited by its fixed 3 × 3 convolution kernels, which can lead to the loss of vessel features at different scales during feature extraction. To address this issue, a segmentation network based on the InceptionNeXt structure is proposed, replacing the 7 × 7 convolution kernel in the ConvNeXt structure with four parallel branches. This modification allows the network to retain performance while achieving higher data throughput. The network adopts the multi-scale feature extraction idea of U-Net and introduces the SE-Net module to realize the inter-channel relationships of extracted features, reducing the weight of unimportant features and improving the accuracy of segmentation results. To evaluate the performance of the proposed model, experiments were conducted on the publicly available DRIVE and CHASE_DB1 datasets by using both pixel-level classification metrics and segmentation-oriented metrics. On the DRIVE dataset, the proposed method achieved sensitivity, specificity, accuracy, AUC, F1-score, IoU and SSIM values of 0.7962, 0.9804, 0.9569, 0.9797, 0.8205, 0.6956 and 0.8653, respectively. On the CHASE_DB1 dataset, the corresponding values were 0.7732, 0.9896, 0.9656, 0.9859, 0.8320, 0.7123 and 0.8397, respectively. The results demonstrate that the model effectively enhances the network’s ability to segment small vessels and improves segmentation performance metrics. Additionally, a comparison with other networks’ segmentation results was performed, showing that the proposed segmentation algorithm outperforms other methods in terms of segmentation performance.</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

InceptionNeXt Unet Retinal Vessel Segmentation Algorithm with Channel Attention Mechanism

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Zhi-Hao Zhang,
  • Jie-Sheng Wang,
  • Shuai-Cheng Qi,
  • Xiao-Tian Wang

摘要

Accurate segmentation of retinal blood vessels in fundus images is beneficial for early diagnosis and prevention of retinal diabetic and cardiovascular diseases. However, accurate segmentation of retinal blood vessels is a challenging task due to issues such as small vessels, vascular lesions and uneven background brightness. With the rapid development of computer vision technology and the rise of deep neural networks, the U-Net has demonstrated good performance in solving retinal vessel segmentation problems. However, the classic U-Net structure is limited by its fixed 3 × 3 convolution kernels, which can lead to the loss of vessel features at different scales during feature extraction. To address this issue, a segmentation network based on the InceptionNeXt structure is proposed, replacing the 7 × 7 convolution kernel in the ConvNeXt structure with four parallel branches. This modification allows the network to retain performance while achieving higher data throughput. The network adopts the multi-scale feature extraction idea of U-Net and introduces the SE-Net module to realize the inter-channel relationships of extracted features, reducing the weight of unimportant features and improving the accuracy of segmentation results. To evaluate the performance of the proposed model, experiments were conducted on the publicly available DRIVE and CHASE_DB1 datasets by using both pixel-level classification metrics and segmentation-oriented metrics. On the DRIVE dataset, the proposed method achieved sensitivity, specificity, accuracy, AUC, F1-score, IoU and SSIM values of 0.7962, 0.9804, 0.9569, 0.9797, 0.8205, 0.6956 and 0.8653, respectively. On the CHASE_DB1 dataset, the corresponding values were 0.7732, 0.9896, 0.9656, 0.9859, 0.8320, 0.7123 and 0.8397, respectively. The results demonstrate that the model effectively enhances the network’s ability to segment small vessels and improves segmentation performance metrics. Additionally, a comparison with other networks’ segmentation results was performed, showing that the proposed segmentation algorithm outperforms other methods in terms of segmentation performance.