基于深度学习分割前列腺多参数MRI图像中的骨质结构  

Segmentation of Bone Structures on Prostate Multiparametric MRI Based on Deep Learning: A Preliminary Study

作　　者：刘想 韩超 张耀峰 张大斗 张晓东[1] 王霄英[1] LIU Xiang;HAN Chao;ZHANG Yaofeng;ZHANG Dadou;ZHANG Xiaodong;WANG Xiaoying(Department of Radiology,Peking University First Hospital,Beijing 100034,China;不详)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]北京赛迈特锐医学科技有限公司,北京100011

出　　处：《中国医学影像学杂志》2021年第8期811-816,821,共7页Chinese Journal of Medical Imaging

摘　　要：目的训练3D U-Net模型,研究自动分割前列腺多参数磁共振成像(mpMRI)扫描所获得的扩散加权成像(DWI和表观扩散系数(ADC)图像中骨质结构的可行性。资料与方法回顾性分析105例前列腺mp MRI患者的图像。手工标注DWI_(高)(b=800 s/mm^(2))、DWI_(低)(b=0 s/mm^(2))和ADC图像上的骨质结构。将不同序列组合作为输入数据训练分割模型,评估输入6个不同序列组合对3D U-Net模型性能的影响。模型评价指标包括定量指标(DICE系数、标签体积)和定性指标(主观评分),模型评价标准为计算测试集中全部序列的达标率,超过80%为符合临床应用需求。结果3D U-Net模型预测DWI图像盆腔骨质的DICE值为0.75(0.70~0.81)~0.81(0.73~0.85),ADC图像为0.79(0.78~0.81)~0.83(0.80~0.85),但不同模型间DICE值差异无统计学意义(H_(DWI高)=2.978,P_(DWI高)>0.05;H_(ADC)=1.140,P_(ADC)>0.05)。不同模型间模型预测与人工标注体积差值差异无统计学意义(H_(DWI高)=2.900,P_(DWI高)>0.05;H_(ADC)=2.236,P_(ADC)>0.05)。定性评分模型1和3在DWI_(高)和ADC图像的分割中达标率最高,均在80%以上。结论使用盆腔DWI_(高)联合ADC序列的3D U-Net模型对前列腺mp MRI盆腔骨质结构分割可达到较高的效能,符合临床应用需求。Purpose To investigative the feasibility of training a 3 D U-Net model for automatic segmentation of bone structures on diffusion weighted imaging(DWI)and apparent diffusion coefficient(ADC)maps of prostate multiparametric MRI(mpMRI).Materials and Methods The mpMRI of 105 patients were retrospectively collected and analyzed.The bone structures were manually annotated on DWIhigh(b=800 s/mm^(2)),DWIlow(b=0 s/mm^(2))and ADC maps.A 3 D U-Net model was trained using different combinations of MRI sequences as input and the efficacies of the six different models were further compared.Model evaluation index were included the quantitative index(DICE coefficient,volume of label)and qualitative index(subjective scores).Results The range of DICE values predicted by 3 D U-Net model was 0.75(0.70-0.81)-0.81(0.73-0.85)on DWI and 0.79(0.78-0.81)-0.83(0.80-0.85)on ADC maps,but there was no significant difference in DICE values among these six different models(HDWI_(high)=2.978,PDWI_(high)>0.05;H_(ADC)=1.140,P_(ADC)>0.05).There was no significant difference in the volume of label among these six models(HDWI_(high)=2.900,PDWI_(high)>0.05;H_(ADC)=2.236,P_(ADC)>0.05).The qualitative scores of model 1 and 3 were achieved the highest(all>80%)in both DWIhigh and ADC maps.Conclusion 3 D U-Net model with a combination input of DWIhigh and ADC can achieve good performance for automatic segmentation of pelvic bone structures on mpMRI,which can meet the needs of clinical application.

关 键 词：人工智能 磁共振成像 扩散加权成像 表观扩散系数 分割 骨盆 

分 类 号：R445.2[医药卫生—影像医学与核医学] R697[医药卫生—诊断学]