基于改进ResNet的直肠癌T分期智能诊断方法  

The Development of An Intelligent Diagnostic Approach for T Staging of Rectal Cancer Using An Enhanced ResNet Model

作　　者：吴越 程军强 李国志 WU Yue;CHENG Junqiang;LI Guozhi(Cancer Hospital Affiliated to Zhengzhou University/Henan Cancer Hospital,Zhengzhou,Henan Province 450006,P.R.China)

机构地区：[1]郑州大学附属肿瘤医院/河南省肿瘤医院,450006 [2]欧亚高科数字技术有限公司 [3]河南中医药大学信息技术学院

出　　处：《临床放射学杂志》2025年第4期730-736,共7页Journal of Clinical Radiology

基　　金：河南省高等学校重点科研项目(编号:22A520032);教育部产学合作协同育人项目(编号:221002627010602);河南省科技攻关项目(编号:232102311052、242102211028)。

摘　　要：目的基于MRI数据,依据端到端的计算机辅助诊断算法,构建改进ResNet(RCSID)的直肠癌T分期智能诊断模型,以期提高直肠癌术前T分期的准确性和效率,为临床制定个性化的有效治疗方案提供数据支持。方法在病历系统搜索2023年6月至2024年6月初次于河南省肿瘤医院进行MRI检查并未经治疗干预的直肠癌患者,搜集其MRI检测结果,构建RCSID预测模型和探究改进ResNet模型在直肠癌T分期的预测价值。结果经临床诊断,155例初次入院检查的直肠癌患者,T1(7例)、T2(27例)、T3(97例)和T4(24例)。在模型预测实验中,对比了不同深度学习模型在直肠癌T分期预测中的表现,主要包括AlexNet模型、VGG16模型、ResNet101模型和RCSID模型,结果显示,在相同实验条件下,RCSID模型预测直肠癌T分期的准确率、精准率、召回率及F1分数均分别优于AlexNet模型、VGG16模型和ResNet101模型,应用RCSID模型对评估直肠癌具体T分期的准确率分别为T1(87.00%)、T2(89.00%)、T3(86.00%)和T4(94.00%),其准确率均分别高于AlexNet模型、VGG16模型和ResNet101模型,经受试者工作特征(ROC)曲线分析结果显示,RCSID模型的ROC曲线下面积(AUC)值高达0.98,分别高于AlexNet(0.72)、VGG16(0.81)和ResNet101(0.87),学习集和测试集Calibration曲线经Delong检验,两者间差异均无统计学意义(P>0.05)。统计4种模型的训练时间和推理时间,按应用时间长短排列为:AlexNet模型<VGG16模型<ResNet101模型<RCSID模型。结论RCSID模型在早期直肠癌T分期智能诊断中获得较高的应用价值,其可行性强,临床上可推广使用.Objective The objective of this study is to develop an enhanced ResNet(RCSID)intelligent diagnosis model for T staging of rectal cancer based on magnetic resonance imaging(MRI)data and end-to-end computer-aided diag-nosis algorithm,aiming to enhance the accuracy and efficiency of preoperative T staging in rectal cancer and provide clinical evidence for personalized and effective treatment planning.Methods The MRI results of rectal cancer patients who un-derwent MRI examination at Henan Cancer Hospital without any treatment intervention were collected from June 2022 to early June 2024.Subsequently,a RCSID prediction model was developed to explore the improved ResNet model's predictive value in T staging for rectal cancer.Results According to clinical diagnosis,a total of 155 patients with rectal cancer were initially examined,including 7 cases classified as T1,27 cases classified as T2,97 cases classified as T3,and 24 cases clas-sified as T4.In the model prediction experiment,we conducted a comparative analysis of various deep learning models for T staging prediction in rectal cancer,namely:AlexNet model,VGG16 model,ResNet101 model,and RCSID model.The find-ings demonstrate that under identical experimental conditions,the RCSID model outperformed the AlexNet model,VGG16 model,and ResNet101 model in terms of accuracy,precision,recall,and F1-Score for predicting the T stage of rectal cancer.The RCSID model demonstrated higher accuracy rates in evaluating specific T stages of rectal cancer compared to the Alex-Net,VGG16,and ResNet101 models.Specifically,the accuracy was 87.00%for T1,89.00%for T2,86.00%for T3,and 94.00%for T4.The ROC curve analysis revealed that the RCSID model exhibited a significantly higher AUC value of 0.98,surpassing those of AlexNet(0.72),VGG16(0.81),and ResNet101(0.87).The calibration curves of both the learning set and the test set were evaluated using the DeLong's test,and no statistically significant difference was observed(P>0.05).Furthermore,when considering training time and inference time

关 键 词：直肠癌 改进ResNet 直肠癌T分期 智能诊断 诊断模型 

分 类 号：R73[医药卫生—肿瘤]