医学
放射治疗
队列
肺炎
放射科
医学物理学
模态(人机交互)
肺癌
放射性肺炎
放射治疗计划
影像引导放射治疗
集合(抽象数据类型)
人工智能
队列研究
选择(遗传算法)
核医学
结果(博弈论)
癌症
辐射剂量
放射治疗师
置信区间
阶段(地层学)
计算机断层摄影术
作者
Zhen Zhang,Haixia Shen,Ji Zhu,Yongling Ji
标识
DOI:10.1016/j.ijpt.2026.101817
摘要
Background and Aims: Clinically deployable decision-support for selecting carbon-ion radiotherapy (CIRT) versus photon radiotherapy in non-small cell lung cancer (NSCLC) is limited.We aimed to develop a multi-modal thoracic radiotherapy foundation model that jointly predicts grade ≥2 radiation pneumonitis (RP) and local control (LC) from routine planning data, and summarizes toxicityefficacy trade-offs to support individualized modality selection.Methods: Planning CT, 3D dose distributions, and structure masks (tumor and total lung), with optional key clinical covariates, were analyzed.A 3D masked-autoencoding backbone was selfsupervised pre-trained on 46842 thoracic cancer patients from four academic centers using radiotherapy-specific CT-dose-structure fusion (Model 1).Model 1 was fine-tuned on 890 photon-treated NSCLC patients to predict RP (grade ≥2) and time-to-event LC (Model 2), and externally validated in an independent cohort (n=339).Model 3 was obtained by continuing fine-tuning from Model 2 on 115 CIRT-treated NSCLC patients (30% training, 70% validation).Performance was benchmarked against DVH/clinical models, radiomics-based machine learning, and 3D deep learning baselines using AUC (RP), C-index (LC), calibration, and bootstrap confidence intervals.Interpretability used transformer-compatible attribution (Grad-CAM saliency) to highlight influential dose and anatomical regions.Results: Across tasks and cohorts, the proposed foundationmodel framework consistently outperformed other approaches.In photon-treated NSCLC, Model 2 achieved strong discrimination for both RP and LC, with AUC and C-index values in the range of 0.75-0.82,and maintained robust performance in the independent external validation cohort with good calibration.After continued finetuning on CIRT-treated NSCLC, Model 3 retained competitive performance on the held-out validation set (AUC=0.71,C-index=0.78).Transformer-compatible attribution maps provided clinically plausible explanations, highlighting influential lung and dose-related regions associated with RP risk and areas relevant to LC prediction.Conclusions: A large-scale, multi-modal thoracic radiotherapy foundation model enabled accurate and well-calibrated prediction of RP and LC across photon and carbon-ion NSCLC cohorts.By transferring representations from multi-center pretraining and sequential fine-tuning, the framework improved generalizability under treatment-domain shift and supported individualized modality selection through toxicity-efficacy trade-off profiling, with interpretable region-level attributions that align with clinical expectations.
科研通智能强力驱动
Strongly Powered by AbleSci AI