基于术前CT人工智能网络识别肾透明细胞癌肉瘤样分化与核分级  

Identification of sarcomatoid feature and nuclear grades of clear cell renal cell carcinoma based on artificial intelligence network of computerized tomography

作　　者：林博涵 邱钱仁顺 陈少豪 许宁[1,2,3] 郑清水 薛学义[1,2] 魏勇 Lin Bohan;Qiu Qianrenshun;Chen Shaohao;Xu Ning;Zheng Qingshui;Xue Xueyi;Wei Yong(Department of Urology,the First Affiliated Hospital of Fujian Medical University,Fuzhou 350005,China;Urology Research Institute,the First Affiliated Hospital of Fujian Medical University,Fuzhou 350005,China;Fujian Key Laboratory of Precision Medicine for Cancer,the First Affiliated Hospital of Fujian Medical University,Fuzhou 350005,China)

机构地区：[1]福建医科大学附属第一医院泌尿外科,福州350005 [2]福建医科大学泌尿外科研究所,福州350005 [3]福建省肿瘤精准诊疗重点实验室,福州350005

出　　处：《中华实验外科杂志》2022年第5期966-969,共4页Chinese Journal of Experimental Surgery

基　　金：2019年福建省科技厅社会发展科技引导性项目(2019Y0018)。

摘　　要：目的构建基于术前CT人工智能网络(AI)识别肾透明细胞癌(ccRCC)肉瘤样分化与核分级,为ccRCC术前精准诊断和诊疗方案制定提供依据。方法分析福建医科大学附属第一医院2015年4月至2018年8月共286例ccRCC患者术前CT影像学及临床病理资料,建立基于术前CT的AI网络模型并探究其在识别ccRCC高低核分级及肉瘤样分化中的作用。采用t检验、受试者工作特征(ROC)曲线和DeLong’test进行统计学分析。结果286例患者中,22例(7.7%)伴肉瘤样分化,病理核分级1级39例(13.6%),2级186例(65.0%),3级43例(15.0%),4级18例(6.4%)。AI预测肉瘤样分化组中肉瘤样分化概率明显高于非肉瘤样分化组[(73.66±21.99)%比(37.41±23.40)%,t=-7.012,P<0.01],ROC曲线下面积(AUC)为0.87(0.80~0.94);两组间肿瘤最大径差异无统计学意义(t=-0.538,P>0.05)。核分级低级别组肿瘤最大径明显小于高级别组[(4.83±3.10)cm比(5.85±3.69)cm,t=-2.166,P<0.05],AI预测低级别组中病理低级别概率明显高于病理高级别组[(74.96±18.52)%比(40.28±0.21)%,t=12.713,P<0.01],AUC=0.88(0.83~0.93),优于肿瘤最大径0.62(0.54~0.70)(DeLong’test,Z=4.563,P<0.01)。结论基于术前CT的AI网络有助于识别ccRCC肉瘤样分化与病理核分级。Objective To construct an artificial intelligence(AI)network based on preoperative computerized tomography(CT)to identify sarcomatoid feature and nuclear grades of clear cell renal cell carcinoma(ccRCC).Methods The preoperative CT imaging and clinicopathological data of 286 patients with ccRCC from April 2015 to August 2018 were analyzed retrospectively.The AI network model based on preoperative CT was established and the role of the model in identifying high and low nuclear grades and sarcomatoid differentiation of ccRCC was tested.The t-test,receiver operating characteristic(ROC)curve and Delong’test were used for statistical analysis.Results Among the all patients,22 cases(7.7%)were accompanied by sarcomatoid differentiation.The pathological nuclear grade was grade 1,39(13.6%);grade 2,186(65.0%);grade 3,43(15.0%);and grade 4,18(6.4%).AI predicted the probability of sarcomatoid feature in sarcomatoid differentiation group was significantly higher than that in non-sarcomatoid differentiation group[(73.66±21.99)%vs.(37.41±23.40)%,t=-7.012,P<0.01],area under the ROC curve(AUC)=0.87(0.80-0.94).There was no significant difference in the maximum tumor diameter between them(t=-0.538,P>0.05).The maximum tumor diameter in the low-grade group was significantly smaller than that in the high-grade group[(4.83±3.10)cm vs.(5.85±3.69)cm,t=-2.166,P<0.05].The probability of AI predicting the low-grade pathology in the low-grade group was significantly higher than that in the high-grade pathology group[(74.96±18.52)%vs.(40.28±0.21)%,t=12.713,P<0.01],AUC=0.88(0.83-0.93),which performed better than that in the maximum tumor diameter of 0.62(0.54-0.70)(Delong’test,Z=4.563,P<0.01).Conclusion The AI network based on preoperative CT is helpful in identifying the sarcomatoid feature and nuclear grades of ccRCC.

关 键 词：人工智能网络 肾透明细胞癌 肉瘤样分化 

分 类 号：R737.11[医药卫生—肿瘤]