| 冯博健,吕巧莉,郭扬中,等.基于多尺度Vision Transformer低剂量CT影像模型在肺结节风险评估中的研究[J].肿瘤学杂志,2025,31(9):793-800. |
| 基于多尺度Vision Transformer低剂量CT影像模型在肺结节风险评估中的研究 |
| A Multi-Scale Vision Transformer Model from Low-Dose CT Scan for Assessing Lung Nodule Risk |
| 投稿时间：2025-06-19 |
| DOI：10.11735/j.issn.1671-170X.2025.09.B007 |
|
 |
| 中文关键词: 肺肿瘤 肺结节 低剂量CT 深度学习 影像组学 风险评估 |
| 英文关键词:lung neoplasms lung nodule low-dose CT deep learning radiomics risk assessment |
| 基金项目:江西省科技厅重点研发计划重点项目（2021BBG71006）；江西省卫生健康委科技计划项目（202410059） |
|
| 摘要点击次数: 271 |
| 全文下载次数: 87 |
| 中文摘要: |
| 摘 要： ［目的］ 构建融合低剂量CT（low-dose CT，LDCT）影像的深度学习与影像组学特征的深度学习模型，提高肺结节风险评估的准确性。［方法］ 回顾性纳入2020年7月至2022年3月期间在江西省肿瘤医院和浙江省肿瘤医院接受LDCT检查并经病理诊断明确结节性质的353例患者。采用YOLOv8模型进行肺结节自动检测，并构建多尺度3D-ResViT肺结节分类模型对结节进行高低危分类。同步提取影像组学特征，多维筛选关键特征嵌入模型实现信息互补融合。［结果］ 3D-ResViT肺结节分类模型进行肺结节风险分层受试者工作特征曲线下面积（area under the curve，AUC ）在内部测试集达到 0.868（95%CI：0.802~0.926），在外部测试集AUC为 0.845（95%CI：0.711~0.950）。融合影像组学特征后的深度学习-影像组学联合模型在内部测试集和外部测试集AUC分别提升至0.890（95%CI：0.826~0.946）和0.865（95%CI：0.706~0.990），集成梯度可视化与SHAP分析进一步验证了模型的有效性。［结论］ 深度学习-影像组学联合模型可有效提升LDCT筛查中肺结节的风险评估准确性，具备较高的临床转化潜力。 |
| 英文摘要: |
| Abstract：［Objective］ To develop and validate a deep learning model that integrates deep learning features from low-dose CT (LDCT) scans with radiomic features to improve the accuracy of lung nodule risk stratification. ［Methods］ In this retrospective study， 353 patients with pathologically confirmed lung nodules admitted in Jiangxi Cancer Hospital or Zhejiang Cancer Hospital from July 2020 to March 2022 were included. The YOLOv8 model was employed for automatic nodule detection. A multi-scale 3D-ResViT lung nodule risk model was constructed to classify nodules into high-risk or low-risk categories. Radiomic features were extracted， and key features were selected through multi-dimensional filtering before being embedded into the model for complementary fusion. ［Results］ The 3D-ResViT lung nodule risk model achieved an area under the curve (AUC) of the receiver operating characteristic of 0.868 (95%CI： 0.802~0.926) in the internal test set and 0.845 (95%CI： 0.711~0.950) in the external test set. The integrated model， combining deep learning and radiomic features， improved the AUC to 0.890(95%CI： 0.826~0.946) in the internal test set and 0.865 (95%CI： 0.706~0.990) in the external test set. Model interpretability was enhanced using Gradient-weighted Class Activation Mapping visualizations and SHapley additive explanations (SHAP) analysis. ［Conclusion］ The proposed deep learning-radiomics combined model effectively improves the accuracy of lung nodule risk assessment in LDCT screening， demonstrating strong potential for clinical translation. |
|
在线阅读
查看全文 查看/发表评论 下载PDF阅读器 |