| 翟欢欢,赵静文,刘翔,石蕴玉,汤显,宋家琳,杨少玲.基于注意力门和空洞空间金字塔池化UNet模型提取肝包膜及评估肝硬化[J].中国医学影像技术,2022,38(9):1385~1390 |
| 基于注意力门和空洞空间金字塔池化UNet模型提取肝包膜及评估肝硬化 |
| Attention gates and atrous spatial pyramidal pooling UNet for extracting liver capsule and evaluating liver cirrhosis |
| 投稿时间：2022-03-24 修订日期：2022-06-15 |
| DOI：10.13929/j.issn.1003-3289.2022.09.023 |
| 中文关键词: 肝硬化 神经网络,计算机 超声检查 肝包膜 注意力机制 空洞卷积 |
| 英文关键词:liver cirrhosis neural networks, computer ultrasonography liver capsule attention mechanism atrous convolution |
| 基金项目: |
|
| 摘要点击次数: 2048 |
| 全文下载次数: 445 |
| 中文摘要: |
| 目的 提出注意力门(AG)和空洞空间金字塔池化(ASPP)UNet模型(AA-UNet)，观察其提取肝脏高频超声图像中肝包膜并用于评估肝硬化的价值。方法 纳入47例肝硬化患者及20例非肝脏疾病患者，于肝脏高频声像图中手工标注肝包膜作为标签图像；将AG及ASPP加入UNet，以之提取声像图中的肝包膜；最后加入视觉几何组(VGG)16全连接层和Softmax分类器，评估有无肝硬化及其程度，即正常及轻度、中度肝硬化。采用AA-UNet提取58例颈动脉斑块患者提取颈动脉高频超声声像图中动脉壁，以验证AA-UNet的鲁棒性。结果 AA-UNet提取的肝脏高频超声声像图中的肝包膜与标签图像相似，其交并比、精确率及F分数均大于DeepLabv3+、UNet、UNet+AG及UNet+ASPP提取结果，且用于提取颈动脉高频超声声像图中的动脉壁的效果亦较佳。AA-UNet用于评估肝脏高频超声声像图中正常及轻、中度肝硬化的准确率分别为90.00%、81.67%及78.33%。结论 AA-UNet用于提取肝脏高频超声声像图中的肝包膜及评估轻度肝硬化的效果较佳。 |
| 英文摘要: |
| Objective To develop the attention gates (AG) and atrous spatial pyramidal pooling (ASPP) UNet (AA-UNet), and to observe its value for extracting liver capsule on high frequency ultrasonograms and evaluating liver cirrhosis. Methods Totally 47 patients with liver cirrhosis and 20 patients without liver diseases were enrolled, and the liver capsules on high frequency ultrasonograms were manually labeled as label images. Then, AG and ASPP were added to UNet to extract liver capsules on high frequency ultrasonography. Finally, full connection layers and Softmax classifier of visual geometry group (VGG) 16 were added to evaluate normal, mild and moderate degree of liver cirrhosis. Meanwhile, carotid artery on high-frequency ultrasonography of 58 patients with carotid plaque were collected, and the walls of carotid arteries were extracted using AA-UNet in order to verify its robustness. Results Liver capsules extracted from high-frequency ultrasonograms with AA-UNet were similar to label images, and the intersection over union, precision and F_score of AA-UNet were all larger than those of DeepLabv3+, UNet, UNet+AG and UNet+ASPP. The effect of AA-UNet for extracting walls of carotid artery on high frequency ultrasonograms remained good. The accuracy of AA-UNet for evaluating normal, mild and moderate degree of liver cirrhosis was 90.00%, 81.67% and 78.33%, respectively. Conclusion AA-UNet was effective for extracting liver capsule on high frequency ultrasonograms and evaluating mild liver cirrhosis. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
