基于前馈注意力ConvNeXt模型分割腹部CT中的胰腺  

Segmentation model of pancreas from abdominal CT based on feedforward attention ConvNeXt

作　　者：张文瀚 王永雄[1] 曾福斌 曹洋森 ZHANG Wenhan;WANG Yongxiong;ZENG Fubin;CAO Yangsen(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Radiation Oncology,the First Affiliated Hospital of Naval Medical University,Shanghai 200433,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093 [2]海军军医大学第一附属医院放疗科,上海200433

出　　处：《中国医学影像技术》2025年第3期466-472,共7页Chinese Journal of Medical Imaging Technology

基　　金：海军军医大学第一附属医院基础医学研究专项(2023PY22)。

摘　　要：目的观察基于融合前馈注意力(FA)ConvNeXt架构模型(SC 2-Net)用于分割腹部CT图中胰腺的价值。方法纳入80名健康成人(数据集1)及68例胰腺病变患者腹部3D CT图(数据集2)。向ConvNeXt网络模型中引入FA机制,在编码器部分引入可缩放卷积模块(SCB)和特征门控(FG)模块,建立并改进ConvNeXt模型,与其他模型(基于Transformer的Swin UNETR、nnFormer、UNETR、TransBTS模型,以及基于ConvNeXt的3D UX-NET模型)对比观察分割胰腺效果,并对加入的模块进行消融实验。结果SC 2-Net模型可准确分割胰腺,其在数据集1的戴斯相似系数(DSC)、95%豪斯多夫距离(HD95)及平均表面距离(MSD)分别为0.92±0.01、(1.08±0.05)mm及(2.12±0.01)mm,SC 2-Net分割胰腺的DSC及HD95均优于其他模型;在数据集2分别为0.82±0.03、(3.35±0.36)mm及(0.87±0.15)mm,均优于其他模型。SC 2-Net在2个数据集中均分割出完整胰腺,而其他模型均存在欠分割或误分割。FA模块加入基础网络后对分割效果产生了显著影响。结论SC 2-Net可提升分割腹部CT图中胰腺的效果。Objective To observe the performance of ConvNeXt architecture model(SC 2-Net)integrated with feedforward attention(FA)for segmentation of pancreas from abdominal CT images.Methods 3D abdominal CT images of 80 healthy adults(Dataset 1)and 68 patients with pancreatic lesions(Dataset 2)were included.ConvNeXt network model was established and enhanced by introducing a FA mechanism,a scalable convolution block(SCB)and a feature gating(FG)module into the encoder section.The performance of the model for segmenting pancreas were comparatively evaluated with other models(Swin UNETR,nnFormer,UNETR,TransBTS models based on Transformer and 3D UX-NET model based on ConvNeXt),while conduct ablation experiments were performed on the added modules.Results SC 2-Net accurately segmented pancreas from abdominal CT images,with Dice similarity coefficient(DSC),95%Hausdorff distance(HD95)and the mean surface distance(MSD)of 0.92±0.01,(1.08±0.05)mm and(2.12±0.01)mm in Dataset 1,respectively.The DSC and HD95 of SC 2-Net segmentation of pancreas were both superior to those of other models.In Dataset 2,SC 2-Net achieved DSC,HD95 and MSD of 0.82±0.03,(3.35±0.36)mm and(0.87±0.15)mm,respectively,surpassing all other models.SC 2-Net achieved complete pancreas segmentation in both datasets,whereas other models demonstrated under-segmentation or mis-segmentation.FA module significantly improved segmentation performance when integrated into the baseline network.Conclusion SC 2-Net could improve segmentation of pancreas from abdominal CT images.

关 键 词：胰腺 体层摄影术 X线计算机 图像分割 

分 类 号：R576[医药卫生—消化系统] R445.3[医药卫生—内科学]