小样本多模态深度学习预测前列腺癌高危生物学特征
回顾性多中心研究纳入377例前列腺癌患者(中心A 341例,中心B 36例),基于18F-PSMA-1007 PET/CT、mpMRI及临床变量构建小样本深度学习模型CL-MGNET,预测ISUP分级、包膜外侵犯及切缘阳性。CL-MGNET预测至少存在一项高危特征时内部测试AUC为0.877,外部验证AUC为0.872,显著优于临床模型的0.792,并优于单一模态模型。
要点
- 融合PET/CT、mpMRI与临床变量的CL-MGNET对高危生物学特征预测AUC达0.872–0.877,优于单模态和临床模型。
- 高危生物学特征定义为ISUP≥3、包膜外侵犯或切缘阳性中至少存在一项。
- 模型在仅30例训练样本下实现跨中心泛化,对外部36例仍保持较高预测效能。
核心结论
- CL-MGNET融合多模态影像与临床变量,在小样本条件下预测前列腺癌高危生物学特征具有较高准确性。
- 该模型预测至少一种高危特征(ISUP≥3、EPE或PSM)的内部测试AUC为0.877,外部验证AUC为0.872。
- 多模态模型显著优于临床模型(AUC 0.792)及PET/CT、mpMRI单一模态模型。
研究设计与样本
- 回顾性多中心研究,共377例,其中中心A 341例,中心B 36例作为独立外部验证队列。
- 输入包括18F-PSMA-1007 PET/CT、mpMRI及临床变量,构建小样本深度学习网络CL-MGNET。
- 模型用30例患者的受限子集训练,并在内部测试集和外部测试集上评估跨中心泛化能力。
诊断标准与关键数值
- 高危侵袭性生物学特征定义为至少存在一项以下特征:ISUP≥3、EPE或PSM。
- CL-MGNET预测高危生物学特征的内部测试AUC为0.877,外部验证AUC为0.872。
- 临床模型预测高危生物学特征的AUC为0.792,CL-MGNET显著优于该模型。
- CL-MGNET在内部和外部测试中均优于PET/CT单模态模型和mpMRI单模态模型。
- 整合临床变量后模型性能显著提升,超过传统方法。
临床应用与阅片要点
- 日常阅片中可将PET/CT与mpMRI关键表现联合临床变量输入该模型,辅助评估前列腺癌高危生物学特征。
- 对外部中心数据仍保持较高AUC,提示可用于数据有限的诊疗环境中辅助风险分层。
- 模型预测的高危特征包括ISUP≥3、包膜外侵犯和切缘阳性,与手术病理决策直接相关。
局限性
- 研究为回顾性设计,外部验证队列仅36例,样本量较小。
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)。
英文原文(摘要)· 点击展开
BACKGROUND: Prostate cancer is the second most common cancer in men, with rising mortality rates necessitating precise risk stratification. High-invasive biological features-specifically International Society of Urological Pathology (ISUP) grade, extracapsular extension (EPE), and positive surgical margins (PSM)-are critical for guiding treatment but are difficult to detect due to tumor heterogeneity. Current imaging modalities, including 18F-PSMA-1007 PET/CT and multiparametric MRI (mpMRI), have limitations in fully capturing these features. This study aims to develop a few-shot deep learning model (CL-MGNET) that integrates multimodal imaging and clinical data to predict high-risk biological features, optimizing performance even with limited training data.
MATERIALS AND METHODS: This retrospective, multicenter study analyzed data from 377 patients: 341 from a primary medical center (Center A) and 36 from an independent external validation cohort (Center B). The study utilized multimodal inputs (PET/CT, mpMRI) and clinical variables to predict ISUP grade, EPE, and PSM. A specialized few-shot deep learning network, CL-MGNET, was designed to fuse these data sources. The model was trained using a restricted subset of 30 patients and subsequently evaluated on both internal and external test sets to assess generalizability across different centers.
RESULTS: CL-MGNET demonstrated excellent performance in predicting high-invasive biological features (defined as the presence of at least one high-risk feature: ISUP ≥ 3, EPE, or PSM), achieving an internal test AUC of 0.877 and an external validation AUC of 0.872, which significantly outperformed the clinical model with an AUC of 0.792. The model surpassed both single-modality models (PET/CT, mpMRI) and the clinical model. Furthermore, CL-MGNET exhibited strong generalization capability, effectively predicting various high-risk biological features. When clinical variables were integrated, the model's performance improved significantly, exceeding traditional methods.
CONCLUSION: The CL-MGNET model, leveraging multimodal imaging data and clinical variables with a few-shot learning approach, successfully predicts high-invasive biological features of prostate cancer with high accuracy, even with limited data. The model's performance across different biological features and medical centers shows its robust generalizability. This method holds great promise for improving prostate cancer diagnosis and risk prediction in data-limited environments.
引用
- Miao, Sun, Yao, et al. Generalizable Deep Learning for Prostate Cancer Risk Stratification: Multicenter Study Integrating 18F-PSMA-1007 PET/CT and mpMRI. Acad Radiol. 2026. PMID: 41547629. doi:10.1016/j.acra.2025.12.050
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。