|本期目录/Table of Contents|

[1]吴倩倩,周蕾蕾,王思雨,等.基于级联卷积神经网络增强CT图像肾脏及肾肿瘤的定位分割研究*[J].生物医学工程研究,2021,04:373-379.
　WU Qianqian,ZHOU Leilei,WANG Siyu,et al.Research on localization and segmentation of kidney and kidney tumors based on cascaded convolutional neural network in enhanced CT images[J].Journal of Biomedical Engineering Research,2021,04:373-379.
点击复制

基于级联卷积神经网络增强CT图像肾脏及肾肿瘤的定位分割研究*(PDF)

《生物医学工程研究》[ISSN:1006-6977/CN:61-1281/TN]

期数:
2021年04期
页码:
373-379
栏目:
出版日期:
2021-12-25

文章信息/Info

Title:
Research on localization and segmentation of kidney and kidney tumors based on cascaded convolutional neural network in enhanced CT images
文章编号:
1672-6278 (2021)04-0373-07
作者:
吴倩倩1; 周蕾蕾2; 王思雨2; 刘成友1; 程龙1; 田书畅1; 蒋红兵1; 3
1. 南京医科大学附属南京医院（南京市第一医院）临床医学工程处，南京 210006；2. 南京医科大学附属南京医院（南京市第一医院）医学影像科，南京 210006；3.南京市急救中心，南京 210003
Author(s):
WU Qianqian1;  ZHOU Leilei2;  WANG Siyu2;  LIU Chengyou1;  CHENG Long1;  TIAN Shuchang1;  JIANG Hongbing1; 3
1. Clinical Medical Engineering Office, Nanjing First Hospital, Nanjing Medical University, Nanjing 210006, China;2. Department of Radiology, Nanjing First Hospital, Nanjing Medical University, Nanjing 210006; 3.Nanjing Emergency Medical Center, Nanjing 210003
关键词:
U-Net; 级联网络; 目标检测; 语义分割; 肾脏; 肾肿瘤; CT图像
Keywords:
U-Net;  Cascaded networks;  Object detection;  Semantic segmentation;  Kidney;  Renal tumor;  CT image
分类号:
R318；TP391；R737.11
DOI:
10.19529/j.cnki.1672-6278.2021.04.06
文献标识码:
A
摘要:
针对目前增强CT图像中肾脏及肾肿瘤分割困难的问题，本研究提出一种基于级联卷积神经网络肾脏及肾肿瘤的定位分割方法。分别以Attention U-Net、Res U-Net、U-Net为基本网络结构，建立级联卷积神经网络模型。一级模型实现正常肾和肿瘤肾的定位，二级模型实现肾脏与肿瘤的精细分割。以影像医师手动勾画为金标准，通过精确度、召回率、F1得分评估定位结果，以Dice、体积相似度（volume similarity ,VS）评估分割精度。由测试结果可见，一级定位模型中，U-Net的整体精确度最高为97.67；最高召回率为90.42，F1得分为90.77，均在Res U-Net中获得。二级分割模型中，级联U-Net的肾脏及肿瘤Dice（0.929）最高，级联Attention U-Net的肾脏（0.895）、肿瘤（0.764）Dice均最高。级联U-Net肾脏及肿瘤、肾脏VS最小（分别为0.085、0.016），肿瘤最优VS（0.009）在级联Attention U-Net中获得。综合全部数据集，级联U-Net模型在定位与分割中表现最佳。基于级联卷积神经网络肾脏及肾肿瘤定位分割模型可在增强CT图像中定位肿瘤肾，且可进一步实现对肾脏及肿瘤的精细分割。
Abstract:
Accurate segmentation of kidney and kidney tumors in enhanced CT images is a difficult problem.We proposed a novel method based on the cascaded convolutional neural network to solve this problem. Attention U-Net, Res U-Net and U-Net were taken as the basic network structure respectively to establish a cascaded convolutional neural network model. The first-level model realized the localization of normal kidney and tumor kidney, the second-level model realized the segmentation of kidney and tumor. Manual outline of radiologists was used as the gold standard to calculate the localization accuracy by Precision, Recall and F1_Score, Dice and volume similarity (VS) were used to evaluate the segmentation accuracy. The test results showed that in the first-level localization model, the overall Precision (97.67) of U-Net was the highest, the highest Recall (90.42)and F1_Score (90.77)were all obtained in the Res U-Net. In the second-level segmentation model, the kidney and tumor Dice (0.929) of the cascaded U-Net was the highest, the kidney (0.895) and the tumor (0.764) Dice of the cascaded Attention U-Net were the highest. The kidney and tumor VS, kidney VS of cascade U-Net were the smallest (they were 0.085 and 0.016 respectively), and the optimal VS of tumor (0.009) was obtained in cascaded Attention U-Net. For KiTS19 and hospital datasets, the cascaded U-Net performed best in localization and segmentation. The cascaded neural network can automatically locate the tumor kidney in the enhanced CT images, and can further realize finer segmentation of the kidney and tumor.

参考文献/References

-

备注/Memo

备注/Memo:
（收稿日期：2021-04-01）南京市医学科技发展资金"青年工程"人才培养专项经费资助项目（QRX11033）；南京医科大学科技发展基金一般项目（NMUB2019155）。△通信作者Email: cmdjhb@126.com
更新日期/Last Update: 2022-01-17