基于深度学习在T2Flair图像分割脑胶质瘤病灶的研究  

Segmentation of glioma on T2Flair images using deep learning model

作　　者：李静[1] 孙静涛[1] 王可欣 张耀峰 李世佳 王霄英[4] LI Jing;SUN Jing-tao;WANG Ke-xin(Tangshan Women and Children’s Hospital,Tangshan063000,China;School of Basic Medical Sciences,Capital Medical University,Beijing100069,China)

机构地区：[1]唐山市妇幼保健院放射科,河北唐山063000 [2]首都医科大学基础医学院,北京100069 [3]北京赛迈特锐医学科技有限公司,北京100069 [4]北京大学第一医院医学影像科,北京100034

出　　处：《中国实验诊断学》2023年第6期639-642,共4页Chinese Journal of Laboratory Diagnosis

基　　金：河北省医学科学研究课题(20221754)。

摘　　要：目的探讨深度学习方法训练模型在颅脑T2Flair图像中自动分割脑胶质瘤的可行性。方法回顾性收集2015年3月到2019年9月脑胶质瘤患者的MRI图像,共获得81位患者的合格T2Flair图像。由1位影像医生逐层标注胶质瘤的范围,由另1位高年资影像医生检查并修改。将标注好的数据按照8∶1∶1的比例随机分为训练集(n=63)、调优集(n=9)和测试集(n=9)用于3D U-Net分割模型的训练。以测试集数据的Dice相似系数(Dice similarity coefficient,DSC)为客观评价指标,评价模型分割效果,并比较模型预测脑胶质瘤体积和医生标注胶质瘤体积的差异。结果测试集中模型预测的DSC值为0.74~0.94,中位数为0.88(0.84,0.90)。医生标注脑胶质瘤的体积为32.7~168.1 cm^(3),中位数为146.0(91.7,162.0)cm^(3),模型预测脑胶质瘤体积为35.8~170.9 cm^(3),中位数为113.0(93.7,134.0)cm^(3),其绝对误差率为0.00~0.23,中位数为0.16(0.07,0.19)。结论基于深度学习模型可初步实现在T2Flair图像中自动分割脑胶质瘤,有望用于脑胶质瘤的智能诊断。Objective To explore the feasibility of deep learning model for automatic segmentation of glioma on T2Flair images.Methods MRI images of 81patients with glioma were collected retrospectively from March 2015to September 2019.The gliomas on T2Flair images were manually labeled by one radiologist and were checked and revised by another radiologist.All data were randomly divided into train set(n=63),validation set(n=9)and test set(n=9)according to the ratio of 8:1:1.The Dice similarity coefficient(DSC)of the test dataset was used to evaluate the model.The volume difference between the manually labeled areas and the predicted areas were calculated.Results The DSCs in the test dataset were 0.74~0.94,and the median was 0.88(0.84,0.90).The volume of the manually labelled areas was 32.7~168.1cm^(3),with a median of 146.0(91.7,162.07)cm^(3).The volume of the predicted labelled areas was 35.8~170.9cm^(3),and the median was 113.0(93.7,134.0)cm^(3).The absolute difference ratio between the manual label and predicted label was 0.00~0.23with a median of 0.16(0.07,0.19).Conclusion The 3DU-Net deep learning model can achieve automatic segmentation glioma on T2Flair images,which make it possible for the automatic diagnosis of glioma in the future.

关 键 词：脑胶质瘤 磁共振成像 深度学习 分割 U-Net 

分 类 号：R-05[医药卫生] R739.41