作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2018, Vol. 44 ›› Issue (12): 258-263,270. doi: 10.19678/j.issn.1000-3428.0050056

• 图形图像处理 • 上一篇    下一篇

基于WRN-PPNet的多模态MRI脑肿瘤全自动分割

朱婷a,b,王瑜a,b,肖洪兵a,b,邢素霞a,b   

  1. 北京工商大学 a.计算机与信息工程学院; b.食品安全大数据技术北京市重点实验室,北京 100048
  • 收稿日期:2018-01-10 出版日期:2018-12-15 发布日期:2018-12-15
  • 作者简介:朱婷(1992—),女,硕士研究生,主研方向为图像处理、模式识别;王瑜(通信作者)、肖洪兵、邢素霞,副教授、博士。
  • 基金资助:

    国家自然科学基金面上项目(61671028);北京市自然科学基金面上项目(4162018);北京市教委社科计划一般项目(KM201510011010)。

Automatic Segmentation of Multimodal MRI Brain Tumors Based on WRN-PPNet

ZHU Tinga,b,WANG Yua,b,XIAO Hongbinga,b,XING Suxiaa,b   

  1. a.School of Computer and Information Engineering; b.Beijing Key Laboratory of Big Data Technology for Food Safety, Beijing Technology and Business University,Beijing 100048,China
  • Received:2018-01-10 Online:2018-12-15 Published:2018-12-15

摘要:

多模态磁共振成像脑肿瘤图像存在灰度不均匀、组织类别多样等缺陷,导致脑肿瘤分割难度大、精度低,且已有脑肿瘤分割算法多为半自动分割算法。为此,建立一种端到端的全自动脑肿瘤分割模型。对脑肿瘤三维图像切片化以获得大量二维切片图像,将训练集的切片图像标准化后直接输入该分割模型,然后用训练好的模型正确分割出脑部神经胶质瘤区域,并采用Dice系数、灵敏度系数以及阳性预测率系数评估模型的分割性能。实验结果表明,该模型操作简单,鲁棒性较好,3个评估指标值分别能够达到0.94、0.92和0.97。

关键词: 多模态磁共振成像, 神经胶质瘤, WRN模块, PPNet模块, 端到端, 全自动分割

Abstract:

Multimodal Magnetic Resonance Imaging(MRI) brain tumors image has many defects,such as uneven gray level,diverse tissue types,which lead to the difficulty and low accuracy of brain tumors segmentation,and most of the existing brain tumors segmentation algorithms are semi-automatic segmentation algorithms.To solve this problem,an end-to-end automatic brain tumors segmentation model is established.A large number of two-dimensional slice images are obtained by slicing the three-dimensional images of brain tumors.The slice images of the training set are normalized and then directly input into the segmentation model.The brain glioma region is correctly segmented by the trained model.The segmentation of the model is evaluated by Dice coefficient,sensitivity coefficient and Positive Predictive Value(PPV) coefficient.Experimental results show that the proposed model is easy to operate and has good robustness,the three evaluation indexes can reach 0.94,0.92 and 0.97 respectively.

Key words: multimodal Magnetic Resonance Imaging(MRI), glioma, WRN module, PPNet module, end-to-end, automatic segmentation

中图分类号: