Liu Xiaohui, Han Xiaowei, Wang Xu, Xu Kaiyuan, Wang Mingliang, Zhang Guozheng
Department of Radiology, The Quzhou Affiliated Hospital of Wenzhou Medical University (Quzhou People's Hospital), Quzhou, China.
Department of Radiology, The Affiliated Cancer Hospital of University of Chinese Academy of Sciences (Zhejiang Cancer Hospital), Hangzhou, China.
Abdom Radiol (NY). 2025 Mar;50(3):1228-1239. doi: 10.1007/s00261-024-04576-2. Epub 2024 Sep 23.
Nuclear grading of clear cell renal cell carcinoma (ccRCC) is crucial for its diagnosis and treatment.
To develop and validate a machine learning model for preoperative assessment of ccRCC nuclear grading using CT radiomics.
This retrospective study analyzed 146 ccRCC patients who underwent surgery between June 2016 and January 2022 at two hospitals (the Quzhou Affiliated Hospital of Wenzhou Medical University with 117 cases and the Affiliated Cancer Hospital of University of Chinese Academy of Sciences with 29 cases). Radiomic features were extracted from preoperative abdominal CT images. Features reduction and selection were carried out using intraclass correlation efficient (ICCs), Spearman rank correlation coefficientsand and the Least Absolute Shrinkage and Selection Operator (LASSO) regression method. Radiomics and clinical models were developed utilizing Support Vector Machine (SVM), Extremely Randomized Trees (Extra Trees), Light Gradient Boosting Machine (LightGBM), Random Forest (RF) and K-Nearest Neighbors (KNN) algorithms. Subsequently, the radiomics nomogramwas developed incorporating independent clinical predictors and Rad_signature. Model performance was evaluated using the area under the curve (AUC), accuracy, sensitivity, and specificity, with decision curve analysis (DCA) assessing its clinical utility.
We extracted 1834 radiomic features from each CT sequence, with 1320 features passing through the ICCs screening process. 480 radiomics features were screened by Spearson correlation coefficient. Then, 15 radiomic features with non-zero coefficient values were determined by Lasso dimensionality reduction technique. The five machine learning methods effectively distinguished nuclear grades. The radiomics nomogram outperformed clinical radiological models and radiomics feature models in predictive performance, with an AUC of 0.936 (95% CI 0.885-0.986) for the training set and 0.896 (95% CI 0.716-1.000) for the external verification set. DCA indicated potential clinical applicability of the nomogram.
The radiomics nomogram, developed by integrating clinically independent risk factors and and Rad_signature, demonstrated robust performance in preoperative ccRCC grading. It offers a non-invasive tool that aids in ccRCC grading and clinical decision-making, with potential to enhance treatment strategies.
透明细胞肾细胞癌（ccRCC）的核分级对其诊断和治疗至关重要。
开发并验证一种基于CT影像组学的机器学习模型，用于ccRCC核分级的术前评估。
这项回顾性研究分析了2016年6月至2022年1月期间在两家医院接受手术的146例ccRCC患者（温州医科大学附属衢州医院117例，中国科学院大学附属肿瘤医院29例）。从术前腹部CT图像中提取影像组学特征。使用组内相关系数（ICCs）、Spearman等级相关系数和最小绝对收缩和选择算子（LASSO）回归方法进行特征降维和选择。利用支持向量机（SVM）、极端随机树（Extra Trees）、轻梯度提升机（LightGBM）、随机森林（RF）和K近邻（KNN）算法建立影像组学和临床模型。随后，结合独立临床预测因子和Rad_signature制定影像组学列线图。使用曲线下面积（AUC）、准确性、敏感性和特异性评估模型性能，并通过决策曲线分析（DCA）评估其临床实用性。
我们从每个CT序列中提取了1834个影像组学特征，其中1320个特征通过了ICCs筛选过程。通过Spearson相关系数筛选出480个影像组学特征。然后，通过Lasso降维技术确定了15个非零系数值的影像组学特征。这五种机器学习方法有效地区分了核分级。影像组学列线图在预测性能方面优于临床放射学模型和影像组学特征模型，训练集的AUC为0.936（95%CI 0.885-0.986），外部验证集的AUC为0.896（95%CI 0.716-1.000）。DCA表明列线图具有潜在的临床适用性。
通过整合临床独立危险因素和Rad_signature开发的影像组学列线图在术前ccRCC分级中表现出强大的性能。它提供了一种非侵入性工具，有助于ccRCC分级和临床决策，有可能改进治疗策略。
