[1]姜梓垚,李翔,魏本征△.多尺度特征融合的膀胱癌磁共振成像分割算法*[J].生物医学工程研究,2024,03:181-189.
JIANG Ziyao,LI Xiang,WEI Benzheng.Segmentation algorithm of bladder cancer MRI image based on multi-scale feature fusion[J].Journal of Biomedical Engineering Research,2024,03:181-189.
点击复制
多尺度特征融合的膀胱癌磁共振成像分割算法*(PDF)
《生物医学工程研究》[ISSN:1006-6977/CN:61-1281/TN]
- 期数:
-
2024年03期
- 页码:
-
181-189
- 栏目:
-
- 出版日期:
-
2024-06-25
文章信息/Info
- Title:
-
Segmentation algorithm of bladder cancer MRI image based on multi-scale feature fusion
- 文章编号:
-
1672-6278 (2024)03-0181-09
- 作者:
-
姜梓垚1; 2; 李翔1; 2; 魏本征1; 2△
-
(1.山东中医药大学 医学人工智能研究中心,青岛 266112; 2. 山东中医药大学 青岛中医药科学院,青岛 266112)
- Author(s):
-
JIANG Ziyao1; 2; LI Xiang1; 2; WEI Benzheng1; 2
-
(1.Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao 266112, China;2.Qingdao Academy of Chinese Medical Sciences, Shandong University of Traditional Chinese Medicine, Qingdao 266112)
-
- 关键词:
-
多尺度特征融合; 膀胱癌; T2加权MRI; Swin Transformer; 对比学习
- Keywords:
-
Multi-scale feature fusion; Bladder cancer; T2-weighted MRI; Swin Transformer; Contrast learning
- 分类号:
-
R318
- DOI:
-
10.19529/j.cnki.1672-6278.2024.03.02
- 文献标识码:
-
A
- 摘要:
-
针对膀胱癌磁共振成像(magnetic resonance imaging,MRI)图像肿瘤区域面积小、膀胱壁边界模糊及像素不平衡等问题,本研究基于特征融合过程和不同图像像素之间的相关性,提出了一种多尺度特征融合的膀胱癌MRI图像分割算法。在编码阶段设计多尺度特征融合模块,用于学习不同编码器的多尺度信息,提取膀胱壁和肿瘤更加丰富的特征;在解码阶段设计的像素对比模块,可增加膀胱壁和膀胱肿瘤间的差异性,解决对比度低及像素不平衡问题,提高膀胱壁与肿瘤相邻边界区域的分割性能,实现膀胱癌多区域分割。本研究在膀胱癌MRI数据集上进行实验,结果显示,算法在膀胱壁和肿瘤区域的Dice分别为89.70%和89.13%、交并比(intersection over union,IoU)分别为81.32%和80.51%、豪斯多夫距离(Hausdorff distance,HD)分别为1.30和1.37,分割结果较已有算法均有一定提升。本研究能较好地辅助临床影像学诊断,可为后续肿瘤分期和临床诊疗提供重要依据。
- Abstract:
-
In view of the small tumor area, blurred bladder wall boundary and pixel imbalance in bladder cancer magnetic resonance imaging (MRI) images, we proposed a multi-scale feature fusion segmentation algorithm for bladder cancer MRI images based on the feature fusion process and the correlation between different image pixels. In the coding stage, a multi-scale feature fusion module was designed to learn the feature information of different encoders, so as to extract richer information of bladder wall and tumor. The pixel contrast module was designed in the decoding stage to increase the difference between bladder wall and bladder tumor, solve the low contrast and pixel imbalance, improve the segmentation performance of the adjacent boundary region between bladder wall and tumor, and realize multi-region segmentation of bladder cancer. The experiments were conducted on the MRI dataset of bladder cancer, the results showed that the Dice of the algorithm in the bladder wall and tumor area was 89.70% and 89.13%, the intersection over union (IoU) was 81.32% and 80.51%, the Hausdorff distance (HD) was 1.30 and 1.37, respectively. Compared with the existing algorithms, the segmentation effect was improved. This research can better assist clinical imaging diagnosis and provide important basis for subsequent tumor staging and clinical diagnosis and treatment.
备注/Memo
- 备注/Memo:
-
(收稿日期:2023-12-05)国家自然科学基金资助项目(62372280,61872225);山东省自然科学基金资助项目(ZR2020KF013,ZR2019ZD04);青岛市科技惠民示范专项项目(23-2-8-smjk-2-nsh)。△通信作者Email: wbz99@sina.com
更新日期/Last Update:
2024-07-18