基于深度学习的肢体骨肉瘤CT医学图像分割  

Deep learning based approach for extremity osteosarcoma segmentation in CT image

作　　者：赵林林[1] 王潜[1] 王军[2] 唐子硕 刘雨 樊卓明 陈继民[1] ZHAO Linlin;WANG Qian;WANG Jun;TANG Zishuo;LIU Yu;FAN Zhuoming;CHEN Jimin(Institute of Laser Engineering,Department of Materials and Manufacturing,Beijing University of Technology,Beijing 100124,China;Bone and Soft Tissue Tumor Treatment Center,Peking University People's Hospital,Beijing 100044,China)

机构地区：[1]北京工业大学材料与制造学部激光工程研究院,北京100124 [2]北京大学人民医院骨与软组织肿瘤治疗中心,北京100044

出　　处：《中国医学物理学杂志》2023年第10期1204-1211,共8页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(82272947);北京大学人民医院研究与发展基金(RDG2021-02,RDL2022-14和RZ2023-02);北京大学临床医学+X青年专项,中央高校基本科研业务费(PKU2023LCXQ016)。

摘　　要：针对骨肉瘤CT图像自动化分割任务,本文制作了骨肉瘤CT图像数据集Osteosarcoma,并提出具有双特征提取结构Double-CNN的D-TransUNet模型。D-TransUNet模型在深度学习TransUNet分割模型的基础上新增特征提取结构。用3×3卷积核进行特征提取,同时在特征通道上进行缩减和拼接,最后将双特征提取结构提取的图像信息进行融合。模型整体提取的原始图像信息更加丰富,进一步提升了分割精度。For the task of automatic segmentation of osteosarcoma in CT images,a dataset of CT images(Osteosarcoma)is established,and a D-TransUNet model with Double-CNN feature extraction structure based on TransUNet is proposed.The 3×3 convolution kernel is used for feature extraction,and the feature channel is reduced and spliced.Finally,the image information extracted by the Double-CNN feature extraction structure is fused.The original image information extracted by the proposed model is more abundant,which further improves the segmentation accuracy.

关 键 词：骨肉瘤 医学图像分割 深度学习 TransUNet模型 D-TransUNet模型 

分 类 号：R318[医药卫生—生物医学工程] R445.3[医药卫生—基础医学]