| 崔新伟,孙尚前,薛志孝.融合nnU-Net与UNeXt的nnUNeXt模型用于分割CT中的肝脏及其内病灶[J].中国介入影像与治疗学,2026,23(4):223-228 |
| 融合nnU-Net与UNeXt的nnUNeXt模型用于分割CT中的肝脏及其内病灶 |
| nnUNeXt model integrated nnU-Net with UNeXt for segmenting liver and its lesions in CT images |
| 投稿时间：2026-03-16 修订日期：2026-04-06 |
| DOI：10.13929/j.issn.1672-8475.2026.04.007 |
| 中文关键词: 肝肿瘤 体层摄影术,X线计算机 轻量化 |
| 英文关键词:liver neoplasms tomography,X-ray computed lightweight |
| 基金项目: |
|
| 摘要点击次数: 11 |
| 全文下载次数: 0 |
| 中文摘要: |
| 目的 观察融合nnU-Net标准化流程与UNeXt轻量化结构的nnUNeXt模型用于分割CT图像中的肝脏及其内病灶的价值。方法 选择肝脏肿瘤分割挑战（LiTS）数据集中125组带标注的肝脏肿瘤CT图像，按4∶1比例划分训练集（n=100）与验证集（n=25）。基于nnU-Net标准化流程、嵌入UNeXt轻量化编码-解码结构构建nnUNeXt模型；基于戴斯相似系数（DSC）、交并比（IoU）、95%豪斯多夫距离（HD95）、参数量及每秒10亿次浮点运算数（GFLOPs），与主流分割模型对比，评价该模型分割肝脏及其内病灶的性能。结果 nnUNet用于分割CT中肝脏的整体性能最佳，其DSC、HD95和IoU分别为0.951±0.024、（1.446±1.721）mm和0.907±0.042。nnUNeXt的DSC、HD95和IoU分别为0.944±0.029、（1.659±2.475）mm及0.896±0.049，与nnUNet较为接近且仅次于nnUNet，用于分割病灶的整体性能最佳。nnUNeXt模型参数量仅为1.472M、GFLOPs仅为26.7，显著低于nnUNet的46.315M和5 565.9；且与原始UNeXt相比，nnUNeXt可在保持相同参数量级基础上使GFLOPs由75.916降至26.7。结论 融合nnU-Net标准化流程与UNeXt轻量化结构的nnUNeXt模型用于分割CT图中的肝脏及其内病灶可在保证分割性能的同时显著降低参数量和计算复杂度。 |
| 英文摘要: |
| Objective To observe the value of nnUNeXt model， which was established through integrating standardized process of nnU-Net with the lightweight architecture of UNeXt， for segmenting liver and its lesions in CT images. Methods Totally 125 annotated CT images of liver tumors from liver tumor segmentation challenge（LiTS） dataset were enrolled and divided into training set （n=100） and validation set （n=25） at the ratio of 4∶1. Based on the standardized process of nnU-Net， the lightweight encoder-decoder structure of UNeXt was embedded to construct nnUNeXt model. The performance of nnUNeXt for segmenting liver and its lesions was evaluated according to Dice similarity coefficient （DSC）， intersection over union （IoU）， 95% Hausdorff distance （HD95）， parameter count and giga floating point operations per second （GFLOPs ） and compared with those of the mainstream segmentation models. Results The overall performance of nnUNet for segmenting liver in CT images was the best， with DSC， HD95 and IoU of 0.951±0.024， （1.446±1.721）mm and 0.907±0.042， respectively. DSC， HD95 and IoU of nnUNeXt was 0.944±0.029，（1.659±2.475）mm and 0.896±0.049， respectively， close to and second only to nnUNet， whicht had the best overall performance for segmenting lesions. The parameter count of nnUNeXt model was only 1.472M， and its GFLOPs was only 26.7， significantly lower than those of nnUNet （46.315M and 5 565.9， respectively）. Compared with the original UNeXt， nnUNeXt could reduce GFLOPs from 75.916 to 26.7 while maintaining the same parameter count level. Conclusion nnUNeXt model， which was established through integrating standardized process of nnU-Net with the lightweight architecture of UNeXt， for segmenting liver and its lesions in CT images could significantly reduce the parameter count and computational complexity while ensuring segmentation performance. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
