深度学习自动生成R.E.N.A.L.肾周径评分并预测临床结局
本研究基于KiTS 2023数据集599例患者训练ResNet-50神经网络,从术前CT自动预测R.E.N.A.L.评分各数值组分(除外前/后定位),并与6名评估者共识评分比较。自动评分与人类共识强相关(Spearman ρ=0.77),对部分/根治性肾切除术、恶性、病理分期≥pT3的预测AUC均高于人类评分。外部验证1806例(193例有人类评分)显示自动评分可预测多种临床结局,有望减少主观性并整合至影像流程。
要点
- 深度学习模型可自动生成R.E.N.A.L.评分各数值组分,与人类共识强相关。
- 自动评分对部分/根治性肾切除术、恶性、分期≥pT3的预测优于个体评估者。
- 外部验证支持自动评分跨机构预测临床结局,减少主观性。
核心结论
- 基于深度学习的肾周径评分可重复性好,与人类评分相关,并能跨机构预测多种临床结局。
- 该方法可减少主观性、简化评估,支持整合至影像工作流程以改善肾癌诊疗。
研究设计与样本
- 使用2023年KiTS Challenge数据集599例患者训练ResNet-50神经网络,输入术前CT图像和专家分割掩膜。
- 采用五折交叉验证生成自动评分,与6名评估者的人类共识评分比较,并分析临床结局关联。
- 外部验证在独立健康系统1806例患者中进行,其中193例有人类评分可用。
主要发现
- 自动评分与人类共识强相关(Spearman ρ=0.77),优于个体评估者(ρ=0.42,P<.01)。
- 自动评分预测部分与根治性肾切除术的AUC为0.87 vs 0.80(P=.0012),恶性为0.72 vs 0.62(P=.0002),病理分期≥pT3为0.81 vs 0.72(P=.0003)。
- 外部队列中,自动评分与人类评分相关,预测根治与部分肾切除术AUC 0.78、更高分期AUC 0.72、高级别病理AUC 0.64、开放手术AUC 0.59。
关键数值与推荐
- 自动评分与人类共识的Spearman相关系数为0.77。
- 预测部分与根治性肾切除术的AUC为0.87 vs 0.80(P=.0012)。
- 预测恶性的AUC为0.72 vs 0.62(P=.0002)。
- 预测病理分期≥pT3的AUC为0.81 vs 0.72(P=.0003)。
临床怎么用
- 可将深度学习自动肾周径评分整合至放射科工作流程,辅助肾癌术前评估。
- 自动评分可减少主观性,为部分与根治性肾切除术决策提供可重复的量化参考。
局限性
- 依赖CT成像。
- 受队列特异性因素影响。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To evaluate whether a deep learning model could automate R.E.N.A.L. nephrometry score generation and predict clinically significant outcomes. METHODS: A ResNet-50 neural network was trained on 599 patients from the 2023 KiTS Challenge dataset to predict numeric R.E.N.A.L. score components (excluding the anterior/posterior designation) using preoperative CT images and expert-derived segmentation masks. Five-fold cross-validation produced automated scores, which were compared with consensus human scores from 6 raters. Associations with clinical outcomes were assessed using logistic regression and receiver operating characteristic analysis. External validation was performed in 1806 patients from an independent health system, with human scores available for 193 cases. RESULTS: Automated scores showed strong correlation with human consensus (Spearman's ρ = 0.77), outperforming individual raters (ρ = 0.42, P <.01). Automated scores demonstrated higher predictive accuracy for partial versus radical nephrectomy (AUC 0.87 vs 0.80, P = .0012), malignancy (AUC 0.72 vs 0.62, P = .0002), and pathologic stage ≥pT3 (AUC 0.81 vs 0.72, P = .0003). In the external cohort, automated scores correlated with human scoring and predicted radical versus partial nephrectomy (AUC 0.78), higher stage disease (AUC 0.72), high-grade pathology (AUC 0.64), and open surgery (AUC 0.59). Limitations include reliance on CT imaging and cohort-specific factors. CONCLUSION: Deep learning-based nephrometry scores are reproducible, correlate with human scoring, and can predict multiple clinical outcomes across institutional cohorts. This approach reduces subjectivity, streamlines assessment, and supports integration into radiology workflows to improve kidney cancer care.
引用
- Wang B, Jonnalagadda R, Siva J, et al. Generation of Automated Nephrometry Scores Through Direct Prediction of Each Component. Urology. 2026. PMID: 41856377. doi:10.1016/j.urology.2026.02.039
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。