用于颈椎MRI分割的多尺度特征融合注意力网络模型  

Multi-Scale Feature Fusion Attention Network Model for Cervical Vertebrae MRI Segmentation

作　　者：周静 钟原 李平[1] 杨毅[2] 马立泰[2] 张涛 ZHOU Jing;ZHONG Yuan;LI Ping;YANG Yi;MA Litai;ZHANG Tao(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China;Department of Orthopedics,West China Hospital of Sichuan University,Chengdu 610041,China)

机构地区：[1]西南石油大学计算机科学学院,成都610500 [2]四川大学华西医院骨科,成都610041

出　　处：《计算机工程》2023年第10期298-304,312,共8页Computer Engineering

基　　金：国家自然科学基金(61873218);西南石油大学创新基金(642)。

摘　　要：近年来,基于深度学习的医学图像辅助诊断逐渐成为主流,但常见的医疗锥体分割模型缺乏对颈椎细节信息的提取,导致锥体分割不完整或边缘相对模糊。为了提高颈椎MRI图像的分割精度,基于ResNet构建一种多尺度特征融合注意力(MSFFA)网络模型。利用多尺度注意力模块融合不同感受野进行注意力特征增强,同时为了降低特征信息融合的损耗,采用跨尺度特征融合模块进行相似域和边缘域特征增强,最终将原始样本的特征信息整合到分割结果中进行细节增强,进一步优化模型分割性能。实验结果表明,MSFFA模型相比于U-Net、AttUNet等模型分割得到的颈椎结构更完整、边缘更平滑,同时在腰椎分割中也能取得更精确的分割结果,并且相比于最优对照模型DeepLabv3+,Dice相似系数的均值提升了1.05个百分点。In recent years,deep learning has been increasingly employed to effectively assist the diagnosis of medical images.However,common medical cone segmentation models cannot extract cervical spine details,resulting in incomplete cone segmentation or relatively blurred edges.To improve the segmentation accuracy of cervical vertebrae Magnetic Resonance Imaging(MRI),a Multi-Scale Feature Fusion Attention(MSFFA)network model based on ResNet is proposed.The multi-scale attention module fuses different receptive fields for attention feature enhancement.Further,to reduce the loss of feature information fusion,a cross-scale feature fusion module is used to enhance the features of similar and edge domains.Finally,the feature information of the original sample is integrated into the segmentation results for detail enhancement,further optimizing the segmentation performance of the model.Experimental results show that compared with U-Net,AttUNet,and other models,the MSFFA model provides a more complete cervical vertebrae structure and smoother edges and can achieve more accurate segmentation results in lumbar vertebrae segmentation.Compared with the best model,DeepLabv3+,the mean Dice Similarity Coefficient(DSC)increases by 1.05 percentage points.

关 键 词：颈椎分割 注意力机制 多尺度融合 特征增强 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]