利用机器学习开发基于CT影像组学的模型以评估分肾功能。

Developing a CT radiomics-based model for assessing split renal function using machine learning.

作者信息

Zhan Yihua, Zheng Junjiong, Chen Xutao, Chen Yushu, Fang Chao, Lai Cong, Dai Mingzhou, Wu Zhikai, Wu Han, Yu Taihui, Huang Jian, Yu Hao

机构信息

The Sun Yat-Sen Memorial Hospital of Sun Yat-Sen University, Guangzhou, China.

The Eighth Affiliated Hospital of Sun Yat-Sen University, Shenzhen, China.

出版信息

Jpn J Radiol. 2025 Jun 3. doi: 10.1007/s11604-025-01786-6.

DOI:10.1007/s11604-025-01786-6

PMID:40459698

Abstract

PURPOSE

This study aims to investigate whether non-contrast computed tomography radiomics can effectively reflect split renal function and to develop a radiomics model for its assessment.

MATERIALS AND METHODS

This retrospective study included kidneys from the study center and split them into training (70%) and testing (30%) sets. Renal dynamic imaging was used as the reference standard for measuring split renal function. Based on chronic kidney disease staging, kidneys were categorized into three groups according to glomerular filtration rate: > 45 ml/min/1.73 m, 30-45 ml/min/1.73 m, and < 30 ml/min/1.73 m.Features were selected based on feature importance ranking from a tree model, and a random forest radiomics model was built.

RESULTS

A total of 543 kidneys were included, with 381 in the training set and 162 in the testing set. In the training set, 16 features identified as most important for distinguishing between the groups were ultimately included to develop the random forest model. The model demonstrated good discriminatory ability in the testing set. The AUC for the > 45 ml/min/1.73 m, 30-45 ml/min/1.73 m, and < 30 ml/min/1.73 m categories were 0.859 (95% CI 0.804-0.910), 0.679 (95% CI 0.589-0.760), and 0.901 (95% CI 0.848-0.946), respectively. The calibration curves for the kidneys in each group closely align with the diagonal, with Hosmer-Lemeshow test P-values of 0.124, 0.241, and 0.199 for the three groups, respectively (all P > 0.05). The decision curve analysis confirmed the radiomics model's clinical utility, demonstrating significantly higher net benefit than both treat-all and treat-none strategies at clinically relevant probability thresholds: 1-69% and 71-75% for the > 45 ml/min/1.73 m group, 15-d50% for the 30-45 ml/min/1.73 m group, and 0-99% for the < 30 ml/min/1.73 m group.

CONCLUSION

Non-contrast computed tomography radiomics can effectively reflect split renal function information, and the model developed based on it can accurately assess split renal function, holding great potential for clinical application.

摘要

目的

本研究旨在探讨非增强计算机断层扫描放射组学能否有效反映分肾功能，并建立用于评估的放射组学模型。

材料与方法

本回顾性研究纳入了研究中心的肾脏，并将其分为训练集（70%）和测试集（30%）。肾动态成像用作测量分肾功能的参考标准。根据慢性肾脏病分期，根据肾小球滤过率将肾脏分为三组：>45 ml/min/1.73m²、30 - 45 ml/min/1.73m²和<30 ml/min/1.73m²。基于树模型的特征重要性排名选择特征，并建立随机森林放射组学模型。

结果

共纳入543个肾脏，其中训练集381个，测试集162个。在训练集中，最终纳入了16个被确定为区分各组最重要的特征来建立随机森林模型。该模型在测试集中显示出良好的鉴别能力。>45 ml/min/1.73m²、30 - 45 ml/min/1.73m²和<30 ml/min/1.73m²类别对应的曲线下面积（AUC）分别为0.859（95%可信区间0.804 - 0.910）、0.679（95%可信区间0.589 - 0.760）和0.901（95%可信区间0.848 - 0.946）。每组肾脏的校准曲线与对角线紧密对齐，三组的Hosmer-Lemeshow检验P值分别为0.124、0.241和0.199（均P>0.05）。决策曲线分析证实了放射组学模型的临床实用性，在临床相关概率阈值下，其净效益显著高于全治疗和不治疗策略：>45 ml/min/1.73m²组为1% - 69%和71% - 75%，30 - 45 ml/min/1.73m²组为15% - 50%，<30 ml/min/1.73m²组为0% - 99%。