王翔,史操,袁正一.融合分区注意力UNet模型用于分割MRI中的膝关节软骨[J].中国医学影像技术,2024,40(5):764~768
融合分区注意力UNet模型用于分割MRI中的膝关节软骨
UNet fusing patch attention for segmenting knee cartilage on MRI
投稿时间:2023-10-27  修订日期:2024-02-26
DOI:10.13929/j.issn.1003-3289.2024.05.027
中文关键词:  膝关节  软骨  深度学习  磁共振成像  注意力机制
英文关键词:knee joint  cartilage  deep learning  magnetic resonance imaging  attention mechanism
基金项目:
作者单位E-mail
王翔 青岛科技大学信息科学技术学院, 山东 青岛 266000  
史操 青岛科技大学信息科学技术学院, 山东 青岛 266000 caoshi@yeah.net 
袁正一 青岛科技大学信息科学技术学院, 山东 青岛 266000  
摘要点击次数: 124
全文下载次数: 112
中文摘要:
      目的 构建融合分区注意力的UNet(PA-UNet)模型,观察其分割MRI中的膝关节软骨的价值。方法 对来源于Osteoarthritis Initiative-Zuse Institute Berlin数据集的膝关节MRI进行切片及预处理,以UNet为骨干网络构建基于分区注意力机制的PA-UNet模型,通过主、客观评价比较该模型与其他模型分割股骨软骨及胫骨软骨的效果;分别以基于UNet、基于SE(第2~4层)的UNet(UNet+SE)、+UNet、++UNet、+++UNet、+UNet+、++UNet++及PA-UNet模型的消融实验观察各模型分割膝关节软骨的效果。结果 PA-UNet可准确分割低难度、中等难度及困难样本中的股骨及胫骨软骨,其分割细小结构效果优于SegNet、UNet及DeepLabv3+模型;其分割股骨软骨及胫骨软骨的戴斯相似系数(DSC)及交并比均高于、而豪斯多夫距离均低于UNet、DeepLabv3+、SA-UNet、RA UNet及SegNet模型。以PA-UNet模型分割股骨软骨及胫骨软骨的DSC分别为88.97%及82.72%,均高于UNet、UNet+SE、+UNet、++UNet、+++UNet、+UNet+及++UNet++模型。结论 PA-UNet可完整分割MRI中的膝关节软骨,尤其对细小结构的分割效果良好。
英文摘要:
      Objective To construct a UNet fusing patch attention (PA-UNet) model, and to observe its value for segmenting knee cartilage on MRI. Methods Slice and preprocessing were performed on knee MRI selected from Osteoarthritis Initiative-Zuse Institute Berlin dataset. Taken UNet as the backbone network, a PA-UNet model was constructed based on patch attention mechanism. The effect of PA-UNet model and other models for segmenting both femoral cartilage and tibial cartilage were compared by subjective and objective evaluations. Ablation experiments based on UNet, UNet based on SE with layers 2—4 (UNet+SE), +UNet, ++UNet, +++UNet, +U-Net+, ++U-Net++ and PA-UNet models were performed to observe the effect of models for segmenting knee cartilage. Results PA-UNet could accurately segment femoral and tibial cartilage in all simple, medium and difficult samples, which had better segmenting effect on small structures than SegNet, UNet and DeepLabv3+ models. The Dice similarity coefficient (DSC) and intersection over union of PA-UNet model for segmenting femoral and tibial cartilage were both higher, while Hausdorff distance of PA-UNet model was lower than those of UNet, DeepLabv3+, SA-UNet, RA UNet and SegNet models. DSC of PA-UNet model for segmenting femoral cartilage and tibial cartilage was 88.97% and 82.72%, respectively, both higher than those of UNet, UNet+SE, +UNet, ++UNet, ++UNet, +UNet+ and ++UNet++ models. Conclusion PA-UNet could segment knee cartilage completely on MRI, especially for small structures.
查看全文  查看/发表评论  下载PDF阅读器