基于3DSEU-Net不确定性循环焦点平均教师的半监督脑肿瘤分割  

Semi-supervised learning for brain tumor segmentation through 3DSEU-Net as uncertaintyaware mean teacher and cyclical focal loss

作　　者：段逸凡 肖洪兵[1] Rahman Md Mostafizur DUAN Yifan;XIAO Hongbing;Rahman Md Mostafizur(School of Artificial Intelligence,Beijing Technology and Business University,Beijing 100048,China)

机构地区：[1]北京工商大学人工智能学院,北京100048

出　　处：《中国医学物理学杂志》2023年第9期1121-1126,共6页Chinese Journal of Medical Physics

基　　金：北京市自然科学基金-北京市教育委员会科技计划重点联合项目(KZ202110011015)。

摘　　要：准确、完整地定位和分割脑肿瘤对脑胶质瘤患者的存活率以及治疗方案的确定起着决定性作用。在三维核磁共振影像(MRI)中,生成准确的注释需要大量的专业知识和时间成本,使用少量有标签数据与大量无标签数据进行半监督学习更加符合实际的临床场景与需求。为此,本文提出一种3DSEU-Net作为半监督模型中的教师与学生网络,该网络引入注意力计算,同时结合跳跃连接,以便获取三维医学影像中更加丰富鲁棒的结构与细节特征,训练过程中,教师模型通过不确定性量化,然后指导学生模型,使学生模型学习到置信度更高的结果,在仅有少量有标签数据的情况下学习到更多的知识,以提升模型的脑肿瘤分割精度。在仅有25个有标签数据的情况下,分割精度比全监督学习提升了12.9%,最高分割精度达81.41%,优于目前可同基准复现的6种半监督方法,证明了本文方法的可行性和有效性。The accurate localization and segmentation of brain tumors greatly affects the survival rate of glioma patients and the determination of treatment schemes.Generating accurate annotations in three-dimensional(3D)magnetic resonance imaging(MRI)requires a lot of professional knowledge and is time-consuming.The semi-supervised learning using a small amount of labeled data and a large amount of unlabeled data is more practical in clinic.Herein a 3DSEU-Net in which squeeze and excitation block is introduced and combined with skip connections is proposed as teacher and student networks in the semi-supervised model,so that the richer and more robust structural and detailed features can be extracted from 3D medical image.During training,the teacher model guides the student model by quantifying uncertainties,which makes the student model learn the results with higher degree of confidence.The proposed model is able to learn more knowledge under the condition that only a small amount of labeled data is available,thereby improving the segmentation accuracy of brain tumors.In the case of only 25 labeled data,the proposed method improves segmentation accuracy by 12.9%over fully supervised learning,and has a highest segmentation accuracy of 81.41%,outperforming 6 semi-supervised methods currently reproducible on the same benchmark.These results verify the feasibility and effectiveness of the proposed method.

关 键 词：三维卷积神经网络 通道注意力 半监督学习 脑肿瘤分割 循环焦点损失 

分 类 号：R318[医药卫生—生物医学工程] TP183[医药卫生—基础医学]