特征区域再聚焦提升全卷积神经网络勾画较小靶区准确度  

Feature area refocusing for improving the accuracy of small target area segmentations by fully convolutional networks

作　　者：蒋家良 罗勇[1] 何奕松 余行 傅玉川[1] JIANG Jialiang;LUO Yong;HE Yisong;YU Hang;FU Yuchuan(Department of Radiotherapy,West China Hospital,Sichuan University,Chengdu 610041,China)

机构地区：[1]四川大学华西医院放疗科

出　　处：《中国医学物理学杂志》2020年第1期75-78,共4页Chinese Journal of Medical Physics

摘　　要：目的:在卷积神经网络基础之上介绍一种特征区域再聚焦的勾画方法。方法:利用端到端的全卷积神经网络,采用正常勾画及特征区域再聚焦勾画两种方法分别对鼻咽癌肿瘤体积(GTVnx)进行自动勾画。选取60例鼻咽癌患者数据,其中40例用于训练,20例用于测试。Dice相似系数(DSC)用于评估自动勾画准确度。结果:正常勾画方法DSC为0.352±0.084,特征区域再聚焦方法DSC为0.524±0.079。对20例测试例勾画结果进行统计学检验结果显示P<0.01。结论:相比正常勾画方法,特征再聚焦勾画方法能够提高对GTVnx的勾画效果,提升较小靶区的勾画精度。Objective To propose a feature area refocusing method based on convolutional neural networks(CNN)for improving the accuracy of small target area segmentations.Methods End-to-end fully convolutional networks(FCN)was used to automatically segment nasopharynx gross tumor volume(GTVnx)by conventional segmentation method and feature area refocusing method.Sixty cases of nasopharyngeal carcinoma were analyzed in the study,with 40 cases for training and 20 cases for testing.Dice similarity coefficient(DSC)was used to evaluate the accuracy of automatic segmentation.Results The DSC obtained by conventional segmentation method was 0.352±0.084,lower than 0.524±0.079 which was obtained by feature area refocusing method.Statistical analysis of 20 test cases showed that the P value was less than 0.01.Conclusion Compared with conventional segmentation method,the feature area refocusing method for segmentation can achieve a better GTVnx segmentation result and improve the accuracy of small target area segmentation.

关 键 词：鼻咽癌 肿瘤体积 靶区勾画 卷积神经网络 注意力机制 

分 类 号：R739.6[医药卫生—肿瘤] R319[医药卫生—临床医学]