机构地区:[1]哈尔滨医科大学附属第二医院影像科,150086
出 处:《临床放射学杂志》2025年第3期507-512,共6页Journal of Clinical Radiology
摘 要:目的观察基于瘤内和瘤周CT图像构建的影像组学模型预测肾透明细胞癌(ccRCC)WHO/ISUP分级的价值,并通过Shapley(Shapley Additive exPlanations)算法对模型进行可视化分析。方法回顾性纳入218例经病理证实的WHO/ISUP分级ccRCC患者,分为低等级组(n=155)和高等级组(n=63)。所有患者术前均接受增强CT扫描,按7∶3的比例随机划分为训练集(n=152)与测试集(n=66),采用3D Slicer软件手动勾画肿瘤感兴趣区(ROI-intra),并环形扩张2 mm、5 mm,获得瘤周感兴趣区(Peri-2 mm、Peri-5 mm),手动调整瘤周勾画区域并整合为瘤内和瘤周的感兴趣体积,提取并筛选影像组学特征后,分别建立瘤内、瘤周及瘤内+瘤周的随机森林模型。评估各个模型的曲线下面积(AUC),以挑选出测试集效能最佳的模型,并观察最佳模型的校准度和净收益,最后,采用Shapley算法对特征进行重要性排序和可视化。结果相比于瘤内单一影像组学模型(训练集AUC=0.821;测试集AUC=0.794),瘤内及瘤周影像组学模型在训练集和测试集中均展现出更高的预测性能,其中瘤内及瘤周2 mm的联合模型预测效能最佳,训练集和测试集中AUC分别为0.906和0.868,决策曲线显示联合模型获得较高的临床净获益。在模型的可解释性分析上,灰度共生矩阵和一阶统计量是WHO/ISUP分级预测模型中贡献度最高的两个特征。结论基于瘤内和瘤周的CT影像组学联合模型在术前可无创预测ccRCC的WHO/ISUP分级。Shapley算法可以为模型提供可解释性,保障了模型的实用性。Objective To investigate the value of radiomics models constructed based on intra-tumoral and peritumoral CT images in predicting WHO/ISUP grading of clear cell renal cell carcinoma(ccRCC),with visualization analysis of the model using the Shapley(Shapley Additive exPlanations)algorithm.Methods A retrospective study included 218 patients with pathologically confirmed WHO/ISUP-graded ccRCC,divided into low-grade(n=155)and high-grade groups(n=63).All patients underwent preoperative enhanced CT scans and were randomly divided into a training set(n=152)and a test set(n=66)in a 7∶3 ratio.Tumor regions of interest(ROI-intra)were manually delineated using 3D Slicer software and then expanded by 2 mm and 5 mm to obtain peritumoral regions of interest(Peri-2 mm,Peri-5 mm).The peritumoral delineation areas were manually adjusted and integrated into the intra-tumoral and peritumoral volumes of interest.After extracting and screening radiomics features,random forest(RF)models for intra-tumoral,peritumoral,and intra-tumoral+peritumoral were established.The area under the receiver operating characteristic curve(AUC)of each model was evaluated to select the best-performing model on the test set,and the calibration and net benefit of the best model were observed.Finally,the Shapley algorithm was used for feature importance ranking and visualization.Results Compared with the single intra-tumoral radiomics model(training set AUC=0.821;test set AUC=0.794),the combined intra-tumoral and peritumoral radiomics models showed higher predictive performance in both the training and test sets.The combined model of intra-tumoral and peritumoral 2 mm had the best predictive performance,with AUCs of 0.906 and 0.868 in the training and test sets,respectively.The decision curve showed that the combined model obtained a higher clinical net benefit.Regarding model interpretability,the gray-level co-occurrence matrix and first-order statistics were the two most contributing features in the WHO/ISUP grading prediction model.Conclusion The combine
关 键 词:肾透明细胞癌 计算机断层摄影 影像组学 Shapley算法 WHO/ISUP分级
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...