设为首页 联系我们 加入收藏

当前位置: 网站首页 期刊分类目录 2025第3期 正文

基于协同训练的直肠肿瘤 MRI 半监督分割方法

作者:时间:2025-07-04点击数:


全文下载: 2.pdf


文章编号:1672-6987(2025)03-0143-09;DOI:10. 16351/j. 1672-6987. 2025. 03. 019


李亚楠 1 ,赵雅坤 2 ,金 鑫 2 ,王明甲 1* (1. 青岛科技大学 自动化与电子工程学院,山东 青岛 266061;2. 青岛科创信达科技有限公司,山东 青岛 266000)


要:直肠肿瘤病灶区域的精确分割可以为肿瘤的临床治疗和预后监测提供重要依据。直

肠肿瘤 MRI 图像结构比较复杂,标签数据获取困难且成本高,为了利用大量无标记数据,本研 究采用半监督图像分割方法。针对直肠癌目标靶区在 MRI 上呈现大小、形状、纹理及边界清 晰度存在个体差异性的问题,本研究在 U-Net 网络的基础上设计了一种新型分割网络 DCCBAM-UNet,并结合使用基于协同训练的半监督方法开展直肠肿瘤 MRI 的分割研究。该方 法通过对双学生-教师模型设置不同初始化的方式进行一致性约束,以集成方法获取高质量的 伪标签,引入蒙特卡罗 Dropout 方法度量伪标签的不确定性,减轻低质量伪标签对分割性能的 影响。在使用 30%的训练数据下,该模型的 DICE 达到了 0. 923 4,Jaccard 达到了 0. 654 2,HD 达到了 12. 035。实验结果表明,该模型在医学图像分割的有效性和泛化性上有一定的性能提 升,能有效解决数据集数量少、病灶区域分割难度大的问题。


关键词:半监督学习;直肠肿瘤分割;协同训练;伪标签;一致性正则化


中图分类号:TP 391. 9 文献标志码:A


引用格式:李亚楠,赵雅坤,金鑫,等 . 基于协同训练的直肠肿瘤 MRI 半监督分割方法[J].

青岛科技大学学报(自然科学版),2025,46(3):143-151.


LI Yanan, ZHAO Yakun, JIN Xin, et al. Co-training based semi-supervised segmentation method for MRI of rectal tumors[J]. Journal of Qingdao University of Science and Technology (Natural Science Edition),2025,46(3):143-151.


Co-Training Based Semi-Supervised Segmentation Method for MRI of Rectal Tumors


LI Yanan1 , ZHAO Yakun2 , JIN Xin2 ,WANG Mingjia1 (1. College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061, China;2. Qingdao Kechuang Xinda Technology Co. , Ltd. , Qingdao 266000, China)


Abstract:Accurate segmentation of rectal tumor lesion regions can provide an important basis for clinical treatment and prognosis monitoring of tumors. Rectal tumor MRI image structure is relatively complex, labeled data acquisition is difficult and costly, in order to utilize a large amount of unlabeled data, this paper adopts a semi-supervised image segmentation method. To solve the problem of individual differences in the size, shape, texture and boundary defini

tion of rectal cancer target areas on MRI, a novel segmentation network DC-CBAM-UNet was designed on the basis of U-Net network, and a semi-supervised method based on collaborative training was used to carry out segmentation research on rectal tumor MRI. By applying consis tent constraints on the ways of setting different initializations on the dual student-teacher model, the method adopts the integrated method of Dropout to measure the uncertainty of the pseudo-tags, and reduces the impact of low-quality pseudo-tags on segmentation performance. Using 30% of the training data, the DICE of the model reached 0. 923 4, the Jaccard reached

0. 654 2, and the HD reached 12. 035. The experimental results show that the model has improved the effectiveness and generalization of medical image segmentation, and can effec tively solve the problem of small number of data sets and difficult segmentation of focal areas.


Key words: semi-supervised learning; rectal tumor segmentation; co-training; pseudolabeling; consistent regularization


收稿日期:2024-05-07

基金项目:国家自然科学基金项目(61971253)

作者简介:李亚楠(1998—),女,硕士研究生 . 通信联系人 .



Copyright © 2011-2017 青岛科技大学学报 (自然科学版)