基于3D超分辨率重组的双参数MRI影像组学模型预测前列腺癌  

Prostate Cancer Prediction Based on a Dual-Parameter MRI Radiomics Model Optimized by 3D Super-Resolution Reconstruction

作　　者：罗童 李慕贞 李爰莹 孙晓川 陈伟彬 LUO Tong;LI Muzhen;LI Yuanying(Department of Medical Imaging,Affiliated Hospital of North China University of Science and Technology,Tangshan,Hebei Province 063000,P.R China)

机构地区：[1]华北理工大学附属医院医学影像中心,唐山063000 [2]华北理工大学人工智能学院

出　　处：《临床放射学杂志》2025年第5期919-925,共7页Journal of Clinical Radiology

基　　金：2023年度唐山市人才资助项目(编号:C202303027)。

摘　　要：目的建立一种3D超分辨率重组双参数MRI(bp-MRI)影像组学特征、临床特征相结合的联合模型,评价其术前预测前列腺癌(PCa)的价值。方法回顾性分析436例疑似PCa而进行MRI检查和病理检查的患者资料。搜集其临床及影像资料,经单、多因素Logistic回归分析筛选出临床特征,采用基于深度学习生成对抗网络(GAN)的3D超分辨率重组技术,将MRI图像空间分辨率提高4倍,提取并筛选DWI(b=800 s/mm2)和ADC图像中的影像组学特征,采用极端随机树分类器构建临床模型、DWI模型、ADC模型、bp-MRI模型(DWI特征+ADC特征)及联合模型(临床特征+DWI特征+ADC特征)。应用受试者工作特征(ROC)曲线、校准曲线、临床决策(DCA)曲线、DeLong检验评估并比较各模型预测PCa的效能。基于最优模型,采用SHAP-Value分析可视化特征在模型中的贡献,并建立列线图。结果结合临床特征(年龄、tPSA)与3D超分辨率重组bp-MRI影像组学特征构建的联合模型,在训练集和测试集中AUC值(0.933,0.932)、准确度(86.2%,87.8%)等方面均表现优异。结合ROC曲线及DeLong检验结果显示,训练集和测试集中联合模型的AUC值较临床模型(0.827,0.868)和ADC模型(0.891,0.816)均有显著的提升(均P<0.05)。在校准曲线中,联合模型预测的概率与实际观察结果紧密吻合,显示出极佳的校准性能。DCA曲线显示联合模型在训练集约0.05~0.95的阈值拥有与DWI模型相当,且优于其他3个模型的净收益,在测试集约0.05~0.75、0.80~0.95阈值内拥有较其他模型更高的临床净收益。联合模型的特征在SHAP-Value图中展现出良好的独立性。结论基于3D超分辨率重组bp-MRI影像组学与临床特征构建的术前PCa预测模型,有助于精准识别PCa,减少不必要的侵入性活检。Objective To establish a combined model integrating 3D super-resolution reconstructed biparametric MRI(bp-MRI)radiomics features and clinical characteristics to evaluate its value in preoperatively predicting prostate cancer(PCa).Methods A retrospective analysis was conducted on 436 patients suspected of having PCa who underwent MRI and pathological examinations.Clinical and imaging data were collected,and clinical characteristics were selected through univariate and multivariate logistic regression analysis.A 3D super-resolution reconstruction technology based on deep learning generative adversarial networks(GAN)was used to enhance the spatial resolution of MRI images by four times,and to extract and screen radiomics features from DWI(b=800 s/mm 2)and ADC images.An ExtraTrees(Extremely Randomized Trees)classifier was employed to construct clinical models,DWI models,ADC models,bp-MRI models(DWI features+ADC features),and combined models(clinical features+DWI,ADC features).The efficacy and clinical net benefit of each model in predicting PCa were evaluated and compared using receiver operating characteristic(ROC)curves,calibration curves,clinical decision(DCA)curves,and DeLong tests.Based on the optimal model,SHAP-Value analysis was used to visualize the contribution of features in the model,and a nomogram was established.Results The combined model,which integrates clinical features(age,tPSA)with 3D super-resolution reconstructed bp-MRI radiomics features,demonstrated excellent performance in the training and testing sets in terms of AUC values(0.933,0.932)and accuracy(86.2%,87.8%).According to ROC curves and DeLong tests,the combined model's AUC values were significantly higher than those of the clinical model(0.827,0.868)and ADC model(0.891,0.816)in both the training and testing sets(P<0.05).In the calibration curves,the probabilities predicted by the combined model closely matched the actual observed results,showing excellent calibration performance.DCA curves indicated that the combined model had a net benefit

关 键 词：前列腺癌 磁共振成像 影像组学 3D超分辨率重组 

分 类 号：R737.25[医药卫生—肿瘤]