Speaker
摘要
目的：针对磁共振引导在线自适应放疗中决策高度依赖临床经验、缺乏客观量化评估的问题，提出一种基于MRI/CT双分支形变向量场影像组学的自适应放疗决策预测方法，旨在为临床提供可靠的辅助决策依据。方法：构建端到端的智能辅助分析架构，深度挖掘形变放射组学特征，结合多模态影像特征库与机器学习算法，建立MRI与CT双分支分类预测模型。采用StratifiedGroupKFold和留一法交叉验证进行患者级评估，通过特征融合（Delta一阶统计特征、形变场特征、PTV边缘特征）提升预测性能。结果：多模态特征融合将AUC从基准0.7059提升至0.9698（提升37%）。MRI分支在外部感兴趣区域结合156维特征集时AUC接近0.97；CT分支在Dose_50区域结合122维特征集时达到全局最优性能，AUC为0.9848，敏感性达1.0。双分支AUC均超过0.96。结论：所提方法可在分次治疗期间为自适应放疗决策提供量化支持，有望减少不必要的计划修改，提升治疗精度与效率，具有良好的临床转化潜力。
Abstract
Objective: To address the lack of objective and quantitative evaluation methods for decision-making in MRI-guided online adaptive radiotherapy (ART), which currently relies heavily on clinical experience, this study proposes a prediction method for ART decision-making based on MRI/CT dual-branch deformation vector field radiomics, aiming to provide reliable decision support for clinicians. Methods: An end-to-end intelligent analysis framework was constructed to deeply mine deformation radiomics features. A dual-branch (MRI/CT) classification prediction model was developed by integrating a multimodal radiomics feature library with machine learning algorithms. Patient-level evaluation was performed using StratifiedGroupKFold and leave-one-out cross-validation. Feature fusion (Delta first-order statistics, deformation field features, and PTV edge features) was employed to enhance predictive performance. Results: Multimodal feature fusion improved the AUC from a baseline of 0.7059 to 0.9698 (a 37% increase). The MRI branch achieved an AUC close to 0.97 using a 156-dimensional feature set in the external region of interest. The CT branch reached the global optimal performance with an AUC of 0.9848 and a sensitivity of 1.0 using a 122-dimensional feature set in the Dose_50 region. Both branches achieved AUC values exceeding 0.96. Conclusion: The proposed method provides quantitative support for ART decisions during fractionated treatment, potentially reducing unnecessary plan modifications and improving treatment accuracy and efficiency, demonstrating strong potential for clinical translation.
| 关键词 | 自适应放疗；形变向量场；影像组学；多模态融合；机器学习 |
|---|---|
| Keywords | Adaptive radiotherapy; Deformation vector field; Radiomics; Multimodal fusion; Machine learning |
