迁移学习在放射性肺炎预测中的预训练模型微调研究

Fine‑tuning of pre‑trained models in transfer learning for prediction of radiation pneumonitis

目的为了更有效地预测胸部放疗后的放射性肺炎,该研究旨在通过迁移学习策略对预训练模型进行微调,以获取适合放射性肺炎预测任务的深度学习特征。方法收集314例2013—2018年在天津市肿瘤医院接受根治性放疗肺癌患者的CT图像和剂量数据。为了提取与放射性肺炎相关的深度学习特征,选取在UCF101视频数据集上预训练的3D ResNet50模型,并对其进行了任务适应的微调。评估策略包括:利用主成分分析方法,对比微调前后特征的分布差异,以验证特征的区分度;使用逻辑回归模型,对比使用微调后的特征和原始预训练特征进行预测的性能,并以曲线下面积(AUC)作为评价指标。结果主成分分析证实,微调后的特征在放射性肺炎的正样本和负样本之间展现出明显的区别。另外,基于微调后的特征的分类模型AUC得分为0.65,而基于原始预训练模型特征的得分为0.58。结论通过迁移学习策略对通用预训练模型进行特定任务的微调,能有效地提取适用于放射性肺炎预测的深度学习特征,并显著提高预测的准确性。

Objective To predict radiation pneumonitis more effectively following thoracic radiotherapy,this study aims to fine‐tune pre‐trained models through transfer learning,to extract deep learning features.Methods Between 2013 and 2018,CT images and dosage data from 314 lung cancer patients undergoing radical radiotherapy in the Tianjin Medical University Cancer Institute and Hospital were collected.To extract deep learning features related to radiation pneumonitis,a 3D ResNet50 model pre‐trained with UCF101 video dataset is applied and fine‐tuned against task adaptation.Evaluation strategies Principal Component Analysis(PCA)is used to contrast the feature distribution differences before and after fine‐tuning,to validate their discriminative capacity.A logistic regression model is applied to compare the predictive performance using features post‐fine‐tuning,and the original pre‐trained features,with AUC(Area Under the Curve)as the evaluation metric.Results PCA confirms the fine‐tuned features show a significant distinction between positive and negative samples of radiation pneumonitis.Moreover,the classification model based on fine‐tuned features achieved an AUC score of 0.65,superior to the score of 0.58 using original pre‐trained model features,indicating enhanced feature performance post‐fine‐tuning.Conclusions The study demonstrates that employing transfer learning strategies to fine‐tune generic pre‐trained models for specific tasks,is feasible in effective deep learning features extraction that is suitable for predicting radiation pneumonitis,thus significantly enhancing predictive accuracy.