| 江慧敏,方立铭,仇书涵,武静.基于增强CT构建多任务改进nnU-Net模型分割原发口腔癌病灶及预测患者无复发生存时间[J].中国医学影像技术,2025,41(9):1568~1572 |
| 基于增强CT构建多任务改进nnU-Net模型分割原发口腔癌病灶及预测患者无复发生存时间 |
| Multi-task improved nnU-Net model based on enhanced CT for segmenting primary oral cancer and predicting patients’ relapse free survival |
| 投稿时间：2025-01-03 修订日期：2025-04-20 |
| DOI：10.13929/j.issn.1003-3289.2025.09.024 |
| 中文关键词: 口腔肿瘤|人工智能|体层摄影术,X线计算机|图像分割|无复发生存 |
| 英文关键词:mouth neoplasms|artificial intelligence|tomography, X-ray computed|image segmentation|relapse free survive |
| 基金项目:皖南医学院重点项目科研基金(WK2023ZZD04)。 |
|
| 摘要点击次数: 99 |
| 全文下载次数: 33 |
| 中文摘要: |
| 目的 观察基于增强CT构建多任务改进nnU-Net模型分割原发口腔癌病灶及预测患者无复发生存时间(RFS)的价值。方法 回顾性分析186例原发口腔癌,基于增强CT构建多任务改进nnU-Net模型以分割肿瘤及预测生存时间:首先以nnU-Net为基线网络实施分割肿瘤预训练,通过改良跳跃连接增强解码器提升识别及分割肿瘤的精度;再以单因素及多因素回归分析筛选与RFS显著相关的临床变量,提取影像组学和深度学习特征,构建生存时间预测模型并对上述模型进行微调。按7∶2∶1比例划分训练集、验证集及测试集,以戴斯相似系数(DSC)评估改进后模型的分割性能,利用一致性指数C-index验证模型预测RFS的效能。结果 多任务改进nnU-Net模型分割原发口腔癌病灶的DSC(0.78)优于3D Inception ResNet(0.65)、3D InceptSENet(0.75)及3D U-Net模型(0.69),预测RFS的C-index(0.798)高于Cox回归模型(0.744)、ICARE模型(0.761)、随机森林模型(0.744)、DeepSurv模型(0.735)、nnU-Net模型(0.760)及放射+nnU-Net模型(0.744),且其分割原发口腔癌的DSC及预测RFS的C-index均优于单纯基线网络(分别为0.653、0.649)、基线网络+多尺度融合模块(0.755、0.752)及基线网络联合临床特征(0.764、0.759)、影像组学特征(0.770、0.764)和临床+影像组学特征(0.773、0.761)。结论 多任务改进nnU-Net模型可有效提高分割原发口腔癌病灶精度及预测患者RFS的准确性。 |
| 英文摘要: |
| Objective To observe the value of multi-task improved nnU-Net model based on enhanced CT for segmenting primary oral cancer and predicting patients’ relapse free survival (RFS). Methods Enhanced CT data of 186 cases of primary oral cancer were retrospectively analyzed, and a multi-task improved nnU-Net model was constructed for tumor segmentation and survival prediction tasks. Pre-training of tumor segmentation was completed with nnU-Net as the baseline network, and the accuracy of recognizing and segmenting tumor was improved by enhancing the decoder through the modified skip connection. Then univariable and multivariable regression analyses were used to select clinical features closely associated with RFS. Radiomics and deep learning features were also extracted to construct a survival prediction model, with fine-tuning of the above model. The training set, validation set and test set were divided at a ratio of 7∶2∶1. Dice similarity coefficient (DSC) was used to evaluate the segmentation performance of the modified model, and the consistency index C-index was used to verify the performance of the improved model for predicting RFS. Results DSC of the multi-task improved nnU-Net model (0.78) for segmenting primary oral cancer was superior to that of 3D Inception ResNet (0.65), 3D InceptSENet (0.75) and 3D U-Net models (0.69), respectively, its C-index for predicting RFS (0.798) was higher than that of Cox regression model (0.744), ICARE model (0.761), random forest model (0.744), DeepSurv model (0.735), nnU-Net model (0.760) and radiology+nnU-Net model (0.744), respectively. DSC for segmenting primary oral cancer and C-index for predicting RFS of multi-task improved nnU-Net model were both superior to those of simple baseline network (0.653 and 0.649), baseline network+multi-scale convolution fusion (0.755 and 0.752), as well as baseline network combined with clinical features (0.764 and 0.759), radiomics features (0.770 and 0.764) and clinical+radiomics features (0.773 and 0.761), respectively. Conclusion Multi-task improved nnU-Net model could be used to effectively improve the accuracy of tumor segmentation and predicting patients' RFS. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
