| 刘辰熙,刘利平,任玉兰,余长敏.基于多尺度特征加权融合与分组门控注意力UNet模型用于分割对比增强T1WI中的子宫内膜癌病灶[J].中国医学影像技术,2025,41(12):2050~2055 |
| 基于多尺度特征加权融合与分组门控注意力UNet模型用于分割对比增强T1WI中的子宫内膜癌病灶 |
| Multi-scale feature weighted fusion and group gated attention UNet model for segmentation of endometrial cancer lesions in contrast-enhanced T1WI |
| 投稿时间：2025-05-08 修订日期：2025-11-10 |
| DOI：10.13929/j.issn.1003-3289.2025.12.024 |
| 中文关键词: 子宫内膜肿瘤 磁共振成像 深度学习 |
| 英文关键词:endometrial neoplasms magnetic resonance imaging deep learning |
| 基金项目: |
|
| 摘要点击次数: 88 |
| 全文下载次数: 19 |
| 中文摘要: |
| 目的 观察基于多尺度特征加权融合(MSWF)与分组门控注意力(GGA)UNet(MGA-UNet)模型用于分割对比增强T1WI(CE-T1WI)中的子宫内膜癌(EC)病灶的价值。方法 回顾性纳入327例EC患者的CE-T1WI数据,按8 ∶ 2比例将其分为训练集(n=261)与验证集(n=66)。以UNet为主干网络,将MSWF模块和通道注意力(SE)模块引入编码器,并将GGA机制引入跳跃连接处,构建MGA-UNet模型,观察其分割CE-T1WI中EC病灶的效能,并与其他主流分割模型进行对比。结果 MGA-UNet模型分割CE-T1WI中的EC病灶的准确率、召回率、交并比及戴斯相似系数分别为96.84%、95.73%、92.37%及88.96%,均优于UNet、UNet++、Attention-UNet、SegNet、Swin-Unet及DeepLabV3+模型。结论 MGA-UNet模型可自动分割CE-T1WI中的EC病灶。 |
| 英文摘要: |
| Objective To observe the value of multi-scale feature weighted fusion (MSWF) and group gated attention (GGA) UNet (MGA-UNet) model for segmentation of endometrial cancer (EC) lesions in contrast-enhanced T1WI (CE-T1WI). Methods CE-T1WI data of 327 EC patients were retrospectively enrolled and divided into training set (n=261) and validation set (n=66) at the ratio of 8∶2. Taken UNet as the backbone network, MSWF module and squeeze-and-excitation (SE) module were introduced into the encoder, and GGA mechanism was introduced into the skip connections, then a MGA-UNet model was constructed. The efficacy of the MGA-UNet model for segmenting EC lesions in CE-T1WI was observed and compared with other mainstream segmentation models. Results The accuracy, Recall, intersection over union and Dice similarity coefficient of MGA-UNet model for segmenting EC lesions in CE-T1WI was 96.84%, 95.73%, 92.37% and 88.96%, respectively, all superior to UNet, UNet++, Attention-UNet, SegNet, Swin-Unet and DeepLabV3+ models. Conclusion MGA-UNet model could be used to automately segment EC lesions in CE-T1WI. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
