|
| SCVi-Net：一种基于混合模型的视网膜血管分割方法 |
| SCVi-Net：a mixed model for retinal vascular segmentation |
| 投稿时间：2024-12-25 |
| DOI：10.3969/j.issn.1005-5630.202412250113 |
| 中文关键词: 视网膜血管分割 U-Net 联合注意力机制 Transformer 空洞卷积 SCVi-Net 医学图像处理 |
| 英文关键词:retinal vascular segmentation U-Net joint attention mechanism transformer atrous spatial pyramid pooling SCVi-Net medical image processing |
| 基金项目:国家自然科学基金(61905144) |
|
| 摘要点击次数: 1009 |
| 全文下载次数: 396 |
| 中文摘要: |
| 现有视网膜血管分割方法通常受限于局部感受野，难以有效捕获全局信息。此外，血管结构在不同尺度下的形态差异较大，使得多尺度特征融合变得困难。为了解决上述问题，提出了一种高效的视网膜血管分割模型SCVi-Net。该模型在U-Net的基础上改进了跳跃连接，引入一个新的空间通道联合注意力模块，通过自适应调整空间和通道权重，增强了特征提取能力。通过在编码器最深层加入视觉Transformer模块，SCVi-Net的全局信息捕获能力得到了提升。空洞空间金字塔池化模块能有效提取多尺度特征，可增强网络的鲁棒性。侧边多尺度融合模块通过融合多个侧边输出，优化了训练过程，从而提升了血管区域的分割精度。为评估模型的优越性，在DRIVE、CHASEDB1和STARE 数据集上进行了对比实验，结果表明，SCVi-Net在复杂视网膜血管图像中具有较好的分割精度和鲁棒性。 |
| 英文摘要: |
| The existing retinal vascular segmentation methods are often limited to the local receptive field. It is difficult to effectively capture the global information, and the performance of vascular structures at different scales varies greatly, which makes it difficult to process features at different scales at the same time. In order to address the problems, an efficient retinal vascular segmentation model SCVi-Net is proposed. Based on U-Net, the model improves the hop connection, introduces a new SCA module, and enhances the feature extraction ability by adaptively adjusting the space and channel weights. By adding a ViT module to the deepest layer of the encoder, SCVi-Net improves the ability to capture global information. The ASPP module effectively extracts multi-scale features and enhances the robustness of the network. The SSF module optimizes the training process by fusing multiple side outputs and improves the segmentation accuracy of the vascular region. In order to evaluate the superiority of the model, comparative experiments are conducted on DRIVE, CHASEDB1 and STARE datasets. The results show that SCVi-Net has high segmentation accuracy and robustness in complex retinal vascular images. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
