基于改进TransBTS的脑肿瘤分割算法研究  

Study on brain tumor segmentation algorithm based on improved TransBTS

作　　者：戴昂 宋亚男[1] 徐荣华[1] 方俞泽 DAI Ang;SONG Yanan;XU Ronghua;FANG Yuze(School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学自动化学院,广东广州510006

出　　处：《实验技术与管理》2023年第8期13-19,39,共8页Experimental Technology and Management

基　　金：广东省科技计划项目(2016A020222012);广东省自然科学基金(2018A030313775);广东工业大学青年基金(18ZK0021.);广东工业大学本科教学工程项目(广工大教字[2019]70号);广东工业大学高水平大学建设研究生教育创新计划项目(2018JGMS-09);广东省本科高校在线开放课程指导委员会研究课题(2022ZXKC143)。

摘　　要：针对现有的脑肿瘤分割网络特征融合不够充分,以及存在的模型复杂的问题,提出一个轻量化的改进Trans BTS的脑肿瘤分割网络。首先,设计了基于大核卷积分解与注意力机制的卷积模块,在减少计算量的同时提高了卷积模块对特征的长距离依赖表征能力;其次,简化了Transformer结构,并设计了多尺度特征作为Transformer的输入,以充分感知全局上下文信息;最后,设计了新的跳跃连接,实现解码过程中对全局深层语义与局部浅层语义信息的充分融合。实验结果表明,该文所提方法在公开数据集BraTS 2021上的多个肿瘤区域(ET,TC,WT)的分割指标与TransBTS相比,Dice评分分别提高了0.55%、1.17%、2.55%,Hausdorff距离降低了1.71、1.68、3.07 mm,并且参数量和计算复杂度分别降低了79.18%和79.88%,平均推理时间减少了1.91 s,更好地实现了分割精度和算法复杂度的平衡。Aiming at the insufficiency of feature fusion in existing brain tumor segmentation networks and the complexity of existing models,a lightweight brain tumor segmentation network with improved TransBTS is proposed.Firstly,a convolutional module based on large kernel convolution decomposition and attention mechanism is designed,which reduces the amount of computation and improves the long distance dependence characterization ability of the convolutional module on features.Secondly,the structure of Transformer is simplified and multi-scale features are designed as the input of Transformer to fully perceive the global context information.Finally,a new skip connection is designed to realize the full integration of global deep semantic information and local shallow semantic information in the decoding process.The experimental results show that,compared with TransBTS,the segmentation index of the proposed method for multiple tumor regions(ET,TC,WT)on the public dataset BraTS 2021 has improved Dice score by 0.55%,1.17%,2.55%,and the Hausdorff distance has decreased by 1.71,1.68,3.07 mm,and the number of parameters and computational complexity have decreased by 79.18%and 79.88%respectively,and the average reasoning time has decreased by 1.91 s.It also achieves a better balance between segmentation accuracy and algorithm complexity.

关 键 词：脑肿瘤分割 多尺度特征融合 TRANSFORMER 卷积分解 注意力机制 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]