利用深度学习进行MR图像腮腺肿瘤病灶分割的可行性研究  

The feasibility study of deep learning in the segmentation of parotid gland tumors using multiparametric MR images

作　　者：王艳[1] 田慧[1] 洪悦[1] 许辉[2] WANG Yan;TIAN Hui;HONG Yue;XU Hui(Department of Radiology and Medical Imaging;Maxillo facial Surgery,the People’s Hospital of Xinjiang Uygur Autonomous Region,Urumqi 830001,China)

机构地区：[1]新疆维吾尔自治区人民医院放射影像中心,新疆乌鲁木齐830001 [2]新疆维吾尔自治区人民医院颌面外科,新疆乌鲁木齐830001

出　　处：《实用放射学杂志》2023年第3期367-370,共4页Journal of Practical Radiology

基　　金：新疆维吾尔自治区自然科学基金项目(2019D01C114)。

摘　　要：目的基于MR T_(2)WI、扩散加权成像(DWI)、增强T_(1)WI多模态图像,利用深度学习方法进行腮腺肿瘤的自动分割.方法回顾性分析经病理证实的39例腮腺肿瘤MR图像,左侧16个病灶,右侧26个病灶,共42个病灶.采用2D U-net网络对腮腺肿瘤常规T_(2)WI、DWI和增强T_(1)WI图像进行病灶区域分割,分别采用单通道(T_(2)WI、增强T_(1)WI)、双通道(DWI)和4通道(T_(2)WI+DWI+增强T_(1)WI)图像作为输入网格进行深度学习模型的训练,评估模型对不同肿瘤亚型的分割效能.结果采用多参数图像组合对肿瘤识别率最高,达89.3%.在多形性腺瘤的分割中,采用T_(2)WI和DWI图像训练的模型均有最好的分割性能.在Warthin瘤的病灶分割中,采用T_(2)WI模型戴斯相似性系数(DSC)值最高,4通道模型95%豪斯多夫距离(HD 95 th)最小.在恶性肿瘤中,采用DWI图像分割性能最好.对其他类别的肿瘤分割中,DWI图像训练的模型DSC值最高,增强T_(1)WI模型HD 95 th最小.结论基于MR T_(2)WI、DWI、增强T_(1)WI多模态图像,利用深度学习方法能有效地实现腮腺肿瘤的自动分割.Objective Deep learning was used to automatically segment parotid gland tumors based on multiparametric T_(2)WI,diffusion weighted imaging(DWI)and enhanced T_(1)WI MR images.Methods MR images of 39 cases of parotid gland tumors con-firmed pathologically were analyzed retrospectively,16 lesions on the left side,26 lesions on the right side,a total of 42 lesions.2D U-net was applied for parotid gland tumors on multiparametric MR images.Single-channel(T_(2)WI,enhanced T_(1)WI),two-channel(DWI)and four-channel(T_(2)WI+DWI+enhanced T_(1)WI)images were used as input network for deep learning model training,and the segmentation efficiency of the model for different tumor subtypes was evaluated.Results 89.3%of tumors could be identified with the combination of multiparameter images.Among pleomorphic adenomas,both models trained with T_(2)WI and DWI images had the best segmentation performance.In the segmentation of Warthin tumor,the dice similarity coefficient(DSC)value of T_(2)WI model was the highest,and the 95%Hausdorff distance(HD 95 th)of 4-channel model was the smallest.In malignant tumors,segmentation performance with DWI images was best.For other types of tumors,the segmentation result had highest DSC with DWI images,and smallest HD 95 th with T_(1)WI images.Conclusion Based on multiparametric T_(2)WI,DWI,and enhanced T_(1)WI MR images,the deep learning methods can be used effectively for the segmentation of parotid gland tumors.

关 键 词：腮腺肿瘤 深度学习 图像分割 磁共振成像 

分 类 号：R739.81[医药卫生—肿瘤] R445.2[医药卫生—临床医学]