基于肾肿瘤病灶部位CT影像自动识别分割方法的设计与实现  

Automated identification and segmentation of renal tumor lesions based on CT imaging

作　　者：王佳蕊 张博文 刘玉洁 刘兆清 张玉婷[3] 岳梦颖 李扬 赵建洪[3] 李书艳 Wang Jiarui;Zhang Bowen;Liu Yujie;Liu Zhaoqing;Zhang Yüting;Yue Mengying;Li Yang;Zhao Jianhong;Li Shuyan(School of Medical Information and Engineering,Xuzhou Medical University,Xuzhou 221004,Jiangsu,China;School of Chemistry and Chemical Engineering,Lanzhou University,Lanzhou 730000,China;Department of Imaging,The Second Hospital of Lanzhou University,Lanzhou 730030,China)

机构地区：[1]徐州医科大学医学信息与工程学院,江苏徐州221004 [2]兰州大学化学化工学院,甘肃兰州730000 [3]兰州大学第二医院影像科,甘肃兰州730030

出　　处：《兰州大学学报（医学版）》2023年第8期47-53,59,共8页Journal of Lanzhou University（Medical Sciences）

基　　金：国家重点研发计划资助项目(2020YFC2006600);甘肃省自然科学基金资助项目(22JR5RA972)。

摘　　要：目的开发基于计算机断层扫描图像和深度学习的高效率且高精确度的肾肿瘤语义分割模型。方法基于注意力机制的U-Net网络架构算法,通过损失函数等相关系数评估模型,选择Patchwise-grid和Patchwise-crop两种训练方法,其中Patchwise-grid是将图像分割成数个网络分析,Patchwise-crop是分析图像中随机裁剪的斑块。使用3折交叉验证方法,在训练集和测试集上,通过训练轮次的增加,得到损失函数、dice系数、dice交叉熵的反映值,最终训练确定使用Patchwise-crop方法。结果构建的自动分割模型的3折交叉验证模型性能dice系数分别为0.8868,0.8726,0.8868;dice交叉熵结果分别为0.1803,0.1902,0.1803,结果优于其他方法。结论使用注意力机制U-Net网络架构算法有效实现了对肾、肾肿瘤区域的自动分割,为肾肿瘤的影像学辅助诊断提供技术支持。Objective To develop a high-efficiency and high-accuracy semantic segmentation model for renal tumors based on computed tomography images and deep learning.Methods An attention-based U-Net network architecture algorithm was employed and the modelevaluated using loss functions and relevant coefficients.Two training methods,Patchwise-grid and Patchwise-crop,were selected and compared.The Patchwise-grid method segmented the image into a network of segments for analysis,while the Patchwise-crop method randomly crops sections of the image for analysis.Three-fold cross-validation was used to evaluate the model performance in the training and testing datasets.The loss function and the dice-soft and dice-crossentropy metrics were employed to assess the model's performance.Results The Patchwise-crop method was the most effective training method,with dice-soft scores of 0.8868,0.8726 and 0.8868,and dice-crossentropy scores of 0.1803,0.1902 and 0.1803.These results were better than those by other methods.Conclusion The attention U-Net network architecture algorithm effectively achieved automatic segmentation of the kidney and renal tumor areas,providing technical support for the imaging-assisted diagnosis of renal tumors.

关 键 词：肾细胞癌 计算机断层扫描 深度学习 注意力机制U-Net 

分 类 号：R541[医药卫生—心血管疾病]