基于深度学习融合模型的算法在宫颈癌术后患者直肠自动分割中的应用研究  

Application of deep learning fusion model algorithm in rectal auto segmentation of postoperatne patients with cervical cancer

作　　者：谷珊珊[1] 吴青南 刘国才[3] 王运来[1] 戴相昆[1] 葛瑞刚[1] 杨微[1] 王秀楠 郭雯 周瑾 鞠忠建[1] GU Shan-shan;WU Qing-nan;LIU Guo-cai;WANG Yun-lai;DAI Xiang-kun;GE Rui-gang;YANG Wei;WANG Xiu-nan;GUO Wen;ZHOU Jin;JU Zhong-jian(Department of Radiation Oncology,the First Medical Center of Chinese PLA General Hospital,Beijing 100089,China;School of Physics Science and Technology,Wuhan University,Wuhan 430072,China;College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;Department of Radiation Oncology,Beijing Shijitan Hospital,Beijing 100038,China)

机构地区：[1]解放军总医院第一医学中心放射治疗科,北京100089 [2]武汉大学物理科学与技术学院,武汉430072 [3]湖南大学电气与信息工程学院,长沙410082 [4]北京世纪坛医院放射治疗科,北京100038

出　　处：《医疗卫生装备》2020年第7期30-35,共6页Chinese Medical Equipment Journal

基　　金：国家自然科学基金项目(61671204)。

摘　　要：目的:针对宫颈癌术后患者直肠形态变化大、难以自动分割的研究难点,运用Dense V-networks网络模型,基于低数量样本进行训练,实现基于CT图像的直肠三维自动分割。方法:融合密集卷积网络(densely connected convolutional networks,DenseNets)与V型网络(V-network)2个网络模型,生成同时具备密集连接和残差结构的Dense V-networks网络模型,以降低需调整的参数数目、减少冗余计算,加快收敛速度,有效地解决训练三维数据时随深度增加出现的梯度消失/爆炸问题;其次选用100例宫颈癌术后患者的盆腔CT图像数据进行标注,其中80例作为训练集,20例作为测试集。将医生手动分割的直肠轮廓作为金标准,采用戴斯相似性系数、豪斯多夫距离、敏感性指数等6个指标分析融合模型的准确性。结果:6个指标平均值分别为0.84(戴斯相似性系数)、2.11 mm(豪斯多夫距离)、0.77(敏感性指数)、0.78(包容性指数)、2.46 mm(平均距离差)和0.71(质心偏差)。融合模型算法分割的直肠轮廓与医生手动分割的直肠轮廓重合度高。结论:在训练样本数量较少情况下,Dense V-networks融合模型算法可实现对形态差异较大的宫颈癌术后患者直肠的高效、精准分割。Objective To automatically segment the rectum on female pelvic 3D CT image based on low-sample training using Dense V-networks fusion network model so as to overcome the difficulty due to morphological changes after cervical cancer surgery.Methods DenseNets and V-network models were combined to generate Dense V-networks model which had dense link and residual structure,so that the Gradient vanishing/explosion problem due to the increased depth during 3D data training could be solved by reducing the number of parameters and redundant calculation and speeding up the convergence speed.Totally 100 postoperative patients with cervical cancer had their pelvic CT images were used for annotation,80 of which were used as training sets and 20 of which were used as test sets.The manually segmented rectal contour by the doctor was used as the gold standard.The accuracy of the fusion model was analyzed using 6 indicators such as Dice similarity coefficient,Hausdorff distance,sensitivity index,inclusive index,mean distance to agreement and deviation of centroid.Results The average values of the six indicators were as follows:Dice similarity coefficient(0.84),Hausdorff distance(2.11 mm),sensitivity index(0.77),inclusive index(0.78),mean distance to agreement(2.46 mm)and deviation of centroid(0.71).The rectal contour segmented by the fusion model algorithm overlapped well with that segmented manually by the physician.Conclusion With a low sample size the Dense V-networks fusion model algorithm can be used for efficient and accurate segmentation of the rectums with large morphological changes after cervical cancer surgery.

关 键 词：深度学习 多模型融合 卷积神经网络 自动分割 直肠 宫颈癌 

分 类 号：R318[医药卫生—生物医学工程]