Ma Chen-Ying, Fu Yi, Liu Lou, Chen Jie, Li Shu-Yue, Zhang Lu, Zhou Ju-Ying
Department of Radiation Oncology, The First Affiliated Hospital of Soochow University, Suzhou, 215006, China.
Radiat Oncol. 2025 Jul 30;20(1):118. doi: 10.1186/s13014-025-02701-z.
This study aimed to develop and validate a multi-temporal magnetic resonance imaging (MRI)-based delta-radiomics model to accurately predict severe acute radiation enteritis risk in patients undergoing total neoadjuvant therapy (TNT) for locally advanced rectal cancer (LARC).
A retrospective analysis was conducted on the data from 92 patients with LARC who received TNT. All patients underwent pelvic MRI at baseline (pre-treatment) and after neoadjuvant radiotherapy (post-RT). Radiomic features of the primary tumor region were extracted from T2-weighted images at both timepoints. Four delta feature strategies were defined (absolute difference, percent change, ratio, and feature fusion) by concatenating pre- and post-RT features. Severe acute radiation enteritis (SARE) was defined as a composite CTCAE-based symptom score of ≥ 3 within the first 2 weeks of radiotherapy. Features were selected via statistical evaluation and least absolute shrinkage and selection operator regression. Support vector machine (SVM) classifiers were trained using baseline, post-RT, delta, and combined radiomic and clinical features. Model performance was evaluated in an independent test set based on the area under the curve (AUC) value and other metrics.
Only the delta-fusion strategy retained stable radiomic features after selection, and outperformed the difference, percent, and ratio definitions in terms of feature stability and model performance. The SVM model, based on combined delta-fusion radiomics and clinical variables, demonstrated the best predictive performance and generalizability. In the independent test cohort, this combined model demonstrated an AUC value of 0.711, sensitivity of 88.9%, and F1-score of 0.696; these values surpassed those of models built with baseline-only or delta difference features.
Integrating multi-temporal radiomic features via delta-fusion with clinical factors markedly improved early prediction of SARE in LARC. The delta-fusion approach outperformed conventional delta calculations, and demonstrated superior predictive performance. This highlights its potential in guiding individualized TNT sequencing and proactive toxicity management.
NA.
本研究旨在开发并验证一种基于多期磁共振成像（MRI）的增量放射组学模型，以准确预测接受局部晚期直肠癌（LARC）全新辅助治疗（TNT）患者的严重急性放射性肠炎风险。
对92例接受TNT的LARC患者的数据进行回顾性分析。所有患者在基线（治疗前）和新辅助放疗后（放疗后）均接受盆腔MRI检查。在两个时间点从T2加权图像中提取原发肿瘤区域的放射组学特征。通过串联放疗前和放疗后的特征定义了四种增量特征策略（绝对差值、百分比变化、比值和特征融合）。严重急性放射性肠炎（SARE）定义为放疗后前2周内基于CTCAE的综合症状评分≥3分。通过统计评估和最小绝对收缩与选择算子回归选择特征。使用基线、放疗后、增量以及联合放射组学和临床特征训练支持向量机（SVM）分类器。基于曲线下面积（AUC）值和其他指标在独立测试集中评估模型性能。
仅增量融合策略在选择后保留了稳定的放射组学特征，并且在特征稳定性和模型性能方面优于差值、百分比和比值定义。基于联合增量融合放射组学和临床变量的SVM模型表现出最佳的预测性能和可推广性。在独立测试队列中，该联合模型的AUC值为0.711，敏感性为88.9%，F1评分为0.696；这些值超过了仅使用基线或增量差值特征构建的模型。
通过增量融合将多期放射组学特征与临床因素相结合，显著改善了LARC中SARE的早期预测。增量融合方法优于传统的增量计算，并表现出卓越的预测性能。这突出了其在指导个体化TNT测序和积极毒性管理方面的潜力。
无。
