|
| 基于双分支注意力网络的青光眼诊断方法 |
| Diagnosis of glaucoma based on dual-branch attention network |
| |
| DOI： |
| 中文关键词: 青光眼 眼底图像 视杯 /视盘分割 双分支注意力网络 多重注意力融合模块 |
| 英文关键词: glaucoma fundus image OC/OD segmentation dual-branch attention network multi-attention fusion module |
| 基金项目:湖北省技术创新专项重大项目( 2022BEC005). |
|
| 摘要点击次数: 417 |
| 全文下载次数: 292 |
| 中文摘要: |
| 通过分割眼底图像的视杯(OC)与视盘(OD)区域并计算二者直径之比得到的杯盘比(CDR)是诊断青光眼的一个重要指标，然而现有视杯 /视盘分割方法的准确度较低，为此提出一种基于双分支注意力网络的青光眼诊断方法。首先，在图像输入主干网络前使用边界到像素方向(BPD)方法增强眼底图像的轮廓信息；其次，在网络编码器部分结合 ConvNeXt的全局交互优势以及 U-Net的局部处理优势，充分提取全局和局部的病理语义信息；最后，在解码器特征重建阶段采用多重注意力融合模块，通过直接和间接映射重组两个编码器和上采样模块提取的平滑和突出特征，深度挖掘目标区域信息，以提高模型对视杯 /视盘区域分割的准确性。在 REFUGE、DRISHTI-GS和 RIM-ONEr3三个具有互补性的临床数据集上进行对比实验，验证了所设计的改进模块在提高眼底图像分割效果上的有效性，而且本文方法可有效平衡 OC和 OD两个目标区域的分割精度，在定量指标和可视化效果上均优于对比方法。 |
| 英文摘要: |
| The cup-to-disc ratio(CDR)，obtained by segmenting the optic cup(OC)and optic disc(OD)regions in fundus images and calculating the ratio of their diameters，is an important indicator for the diagnosis of glaucoma. However，the existing OC/OD segmentation methods have low accuracy，so a new glaucoma diagnosis algorithm based on dual-branch attention network was proposed. Firstly，the Boundary to Pixel Direction(BPD)method was used to enhance the contour information of fundus images before inputting them into the backbone network. Secondly，in the network encoder part，the global interaction advantage of ConvNeXt was combined with the local processing advantage of U-Net to fully extract both global and local pathological semantic information. Finally，in the feature reconstruction stage of the decoder，a multi-attention fusion module was adopted to deeply excavate the target region information via directly and indirectly mapping and restructuring the smooth and prominent features extracted by two encoders and up-sampling modules，so as to improve the accuracy of OC/OD segmenting.Comparative experiments were conducted on three complementary clinical data sets，i.e. REFUGE，DRISHTI-GS and RIM-ONE r3，which verified the effectiveness of the designed modules in improving the segmentation performance of fundus images. Moreover，the proposed method an effectively balance the segmentation accuracy of OC and OD target regions，outperforming the comparison methods in both quantitative metrics and visual effects. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
