MRI序列以及注意力门、残差网络对U-Net脑肿瘤分割模型的影响  

Effects of MRI sequence and attention gates,residual networks on U-Net brain tumor segmentation model

作　　者：张巨 朱文珍[1] 张顺 朱虹全 吴迪 刘栋 ZHANG Ju;ZHU Wen-zhen;ZHANG Shun(Department of Radiology,Tongji Hospital,Tongji Medical College,Huazhong University of Science and Technology,Wuhan 430030,China)

机构地区：[1]华中科技大学同济医学院附属同济医院放射科,武汉430030

出　　处：《放射学实践》2023年第7期825-829,共5页Radiologic Practice

基　　金：国家自然科学基金联合基金重点支持项目(U22A20354)。

摘　　要：目的:探讨MRI序列选择以及注意力门、残差网络对U-Net脑肿瘤分割模型准确性的影响。方法:使用BraTS 2021的脑肿瘤病例1251例,排除经裁剪后缺乏肿瘤图像特征的病例,以7:2:1的比例分为训练组、验证组和测试组,使用7种不同序列组合(所使用序列包括T_(1)WI、T_(2)WI、T_(2)FLAIR、T_(1)增强)训练U-Net模型,使用Friedman检验和成对比较(经Bonferroni校正法调整显著性值)对比测试集的Dice系数。使用4通道图像以相同方法对比基线U-Net、注意力U-Net、残差U-Net以及注意力残差U-Net对脑肿瘤分割的准确性。结果:在分割肿瘤强化区域、核心区域时,无T_(1)增强序列的3通道组的Dice系数显著低于其他组;在分割全肿瘤时,无T_(2)及T_(2)FLAIR的2通道组的Dice系数显著低于其他组,无T_(2)FLAIR的3通道组显著低于剩余其他组,4通道组及无T_(1)的3通道组显著高于其他组,余组间差异无统计学意义。4种U-Net模型仅在分割全肿瘤时存在显著差异,在进一步的成对比较中差异无统计学意义。结论:MRI序列对U-Net分割表现的影响可能与标注方式、该序列所包含的特征信息等有关。本研究中:单独剔除T_(1)序列对U-Net模型无显著影响;与注意力门相比,残差网络可能一定程度提高了U-Net模型的分割准确度。Objective:To investigate the impact of MRI sequence selection and attention gates,residual networks on the accuracy of the U-Net brain tumor segmentation model.Methods:A total of 1251 brain tumor cases from the BraTS 2021 dataset were used,excluding cases lacking tumor image features after cropping.The cases were divided into training,validation,and test groups in a ratio of 7:2:1.Seven different sequence combinations(including T_(1)WI,T_(2)WI,T_(2) FLAIR,and T_(1) contrast-enhanced)were used to train the U-Net model.The Friedman test and pairwise comparison(significance values were adjusted by the Bonferroni correction)were used to compare the Dice coefficients of the test set.The baseline U-Net,attention U-Net,residual U-Net,and attention-residual U-Net were compared using four-channel images in the same way.Results:The Dice coefficients of the 3-channel group without T_(1)-enhancement sequence were significantly lower than those of the other groups when segmenting the Gd-enhancing tumor and tumor core.When segmenting the whole tumor,the Dice coefficients of the 2-channel group without T_(2) and T_(2) FLAIR were significantly lower than those of the other groups.The Dice coefficients of the 3-channel group without T_(2) FLAIR were significantly lower than those of the other remaining groups,while the 4-channel group and the 3-channel group without T_(1) were significantly higher than the other groups.There was no significant difference between the other groups.There were significant differences among the four U-Net models only when segmenting the whole tumor.There was no significant difference in the corrected pairwise comparisons.Conclu-sion:The impact of MRI sequences on the segmentation performance of the U-Net model may be related to the annotation protocol and the feature information contained in the sequence.In this study,the exclusion of the T_(1) sequence alone did not have a significant impact on the U-Net model.Compared to attention gates,residual networks may have improved the segmentation accuracy o

关 键 词：图像分割 注意力门 残差网络 脑肿瘤 

分 类 号：R445.2[医药卫生—影像医学与核医学] R739.41[医药卫生—诊断学]