基于形状引导和不确定性估计的半监督三维医学图像分割  

Semi-Supervised 3D Medical Image Segmentation Based on Shape Guidance and Uncertainty Estimation

作　　者：宋文彪 许叶彤 王毅[1,2] 杜晓刚 雷涛[1,2] SONG Wenbiao;XU Yetong;WANG Yi;DU Xiaogang;LEI Tao(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi'an,Shaanxi 710021,China;Shaanxi Joint Laboratory of Artificial Intelligence,Shaanxi University of Science and Technology,Xi'an,Shaanxi 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021 [2]陕西科技大学陕西省人工智能联合实验室,陕西西安710021

出　　处：《计算技术与自动化》2024年第4期110-116,共7页Computing Technology and Automation

基　　金：国家自然科学基金资助项目(62271296,62201452,62201334)。

摘　　要：在基于深度学习方法的医学图像分割任务中,通常需要大量的标记数据。然而,获得可靠的标注是昂贵且耗时的。为此,提出了一种新的框架,采用具有形状约束和不确定性估计的双一致性正则化半监督方法,用于3D医学图像分割。首先,引入了一种基于学习目标区域的形状约束,通过联合学习两个网络的输出,加强几何形状约束,从而学习更可靠的信息。其次,设计了一种分割网络,以生成不同尺度的特征图,并引入了多尺度一致性损失来增强其稳定性。然而,由于这些特征图的空间分辨率不同,直接在每个像素上强制一致性可能导致不可靠的结果和信息丢失。因此,进一步提出了一种基于不确定性估计的多尺度一致性学习,以逐步学习有意义和可靠的特征区域,并增强模型的鲁棒性。实验结果表明,由于强大的无标记数据的知识挖掘能力,本文所提出的方法优于流行的半监督医学图像分割方法。In medical image segmentation tasks based on deep learning methods,a large amount of labeled data is often required.However,obtaining reliable annotations is expensive and time-consuming.To solve the above problems,a new framework is proposed for 3D medical image segmentation using a biconsistent regularized semi-supervised method with shape constraints and uncertainty estimation.Firstly,a shape constraint based on the learning target region is introduced,and the geometric constraint is strengthened by jointly learning the output of the two networks,so as to learn more reliable information.Secondly,a segmentation network is designed to generate feature maps at different scales,and multi-scale consistency loss is introduced to enhance its stability.However,due to the different spatial resolutions of these feature maps,forcing consistency directly on each pixel can lead to unreliable results and loss of information.Therefore,a multi-scale consistent learning based on uncertainty estimation is further proposed to gradually learn meaningful and reliable feature regions and enhance the robustness of the model.Experimental results show that the proposed method is superior to the popular semi-supervised medical image segmentation method due to the powerful knowledge mining ability of label-free data.

关 键 词：3D医学图像分割 半监督学习 形状约束 不确定性估计 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]