
浏览全部资源
扫码关注微信
1.江苏大学计算机科学与通信工程学院,镇江 212000
2.江苏大学附属医院影像科,镇江 212001
管昌杰,Email:2222308013@stmail.ujs.edu.cn,研究方向:基于人工智能的医学图像处理。
朱海涛,Email:haitaozhu@ujs.edu.cn。
收稿日期:2025-04-08,
网络出版日期:2025-09-29,
移动端阅览
管昌杰,宋雪桦,张时畅等.MS-UNet:基于改进U-Net的心脏磁共振图像高精度分割方法[J].重庆医科大学学报,
Guan Changjie,Song Xuehua,Zhang Shichang,et al.MS-UNet: a high-precision segmentation method for cardiac magnetic resonance images based on Convolutional Networks for Biomedical Image Segmentation[J].Journal of Chongqing Medical University,
管昌杰,宋雪桦,张时畅等.MS-UNet:基于改进U-Net的心脏磁共振图像高精度分割方法[J].重庆医科大学学报, DOI:10.13406/j.cnki.cyxb..
Guan Changjie,Song Xuehua,Zhang Shichang,et al.MS-UNet: a high-precision segmentation method for cardiac magnetic resonance images based on Convolutional Networks for Biomedical Image Segmentation[J].Journal of Chongqing Medical University, DOI:10.13406/j.cnki.cyxb..
目的
2
基于U-Net网络(convolutional networks for biomedical image segmentation,U-Net)构建一种多尺度特征融合与空间通道压缩激励的MS-UNet网络(multi-scale feature fusion and spatial-channel squeeze-and-excitation-Unet,MS-UNet),以提升心脏磁共振图像中的心室及心肌组织分割性能。
方法
2
通过以下步骤构建MS-UNet网络:①增加下采样与上采样层数;②设计多尺度残差特征融合模块(multi-scale feature fusion,MSRF),替换传统3×3卷积层,融合不同尺度上下文信息;③在上采样与跳跃连接之间引入空间和通道压缩与激励模块(spatial-channel squeeze-and-excitation,scSE),强化关键区域特征权重。基于自动心脏诊断挑战数据集(automated cardiac diagnosis challenge dataset,ACDC),训练并测试MS-UNet的心脏分割性能,并与主流分割网络进行对比。分割性能差异比较采用单因素方差分析、Turkey检验或配对
t
检验。
结果
2
MS-UNet在左心室(left ventricle,LV)、右心室(right ventricle,RV)及心肌(Myoglobin,MYO)分割任务中表现优异:平均相似系数(dice similarity coefficient,Dice)、平均交并比(intersection over union,IOU)、平均H
ausdorff距离(hausdorff distance,HD)均显著优于其他模型,且与多个模型比较的差异均具有统计学意义(均
P
<
0.05)。在不同期相心脏分割性能测评中,除了MYO分割Dice和IOU稍差于收缩末期(End Systolic,ES),舒张末期(End Diastolic,ED)心脏各亚结构分割性能评价指标均最佳。2个期相间各指标差异具有统计学意义(均
P
<
0.05)。消融实验结果表明,网络模型在各心脏亚结构分割中的Dice、IOU随着模块数目增加而逐渐增加,HD逐渐降低,各模型间所有性能评价指标间差异具有统计学意义(均
P
<
0.001)。
结论
2
MS-UNet通过多尺度特征融合与自适应注意力机制,明显提升了心脏磁共振图像分割精度,具有广阔的临床应用前景。
Objective
2
To construct a Multi-Scale Feature Fusion and Spatial-Channel Squeeze-and-Excitation-Unet(MS-UNet) network based on Convolutional Networks for Biomedical Image Segmentation(U-Net),and to improve the performance of ventricular and myocardial tissue segmentation in cardiac MRI images.
Methods
2
The MS-UNet network was constructed by the following steps:(1) increase the number of downsampling and upsampling layers;(2)design a Multi-Scale Residual Feature Fusion(MSRF) module to replace the traditional 3 × 3 convolutional layers and fuse different scales of contextual information;and(3) introduce a Spatial-Channel Squeeze-and-Excitation (scSE) module between the upsampling and the hopping connection to strengthen the feature weights in critical regions. The cardiac segmentation performance of MS-UNet was trained and tested based on the Automated Cardiac Diagnosis Challenge dataset,and it was compared with mainstream segmentation networks. A one-way analysis of variance,the Turkey test,or the paired t-test was used for comparison of segmentation performance.
Results
2
MS-UNet demonstrates outstanding performance in segmentation tasks for the left ventricle(LV),right ventricle(RV),and myocardial tissue(MYO) the mean similarity coefficient(Dice),intersection over union(IOU),and mean Hausdorff distance (HD) were significantly better than other models,and the differences compared with multiple models were statistically significant (all
P
<
0.05). In phase-specific cardiac segmentation performance evaluations,except for slightly lower Dice and IOU in MYO segmentation compared to end systolic(ES),end diastolic(ED) demonstrated optimal segmentation metrics for all cardiac subsystem evaluations. Differences between the two phases were statistically significant(all
P
<
0.05). The results of the ablation experiment indicated that the Dice and IOU of the network model in the segmentation of each cardiac substructure gradually increased with the increase in the number of modules,while HD gradually decreased. The differences in all performance evaluation indicators among the models were statistically significant(all
P
<
0.001).
Conclusion
2
MS-UNet obviously improves the segmentation accuracy of cardiac magnetic resonance images through multi-scale feature fusion and adaptive attention mechanism,and therefore,it has broad clinical application prospects.
Karna V V R , Karna V R , Janamala V , et al . A comprehensive review on heart disease risk prediction using machine learning and deep learning algorithms [J ] . Archives of Computational Methods in Engineering , 2025 , 32 ( 3 ): 1763 - 1795 .
Long J , Shelhamer E , Darrell T . Fully convolutional networks for semantic segmentation [C ] // Proceedings of the IEEE conference on computer vision and pattern recognition . 2015 : 3431 - 3440 .
Zhang H , Zhang W , Shen W , et al . Automatic segmentation of the cardiac MR images based on nested fully convolutional dense network with dilated convolution [J ] . Biomedical signal processing and control , 2021 , 68 : 102684 .
Zhao L , Zhou D , Jin X , et al . Nn-TransUNet:an automatic deep learning pipeline for heart MRI segmentation [J ] . Life , 2022 , 12 ( 10 ): 1570 .
Simantiris G , Tziritas G . Cardiac MRI segmentation with a dilated CNN incorporating domain-specific constraints [J ] . IEEE Journal of Selected Topics in Signal Processing , 2020 , 14 ( 6 ): 1235 - 1243 .
da Silva I F S , Silva A C , de Paiva A C , et al . A cascade approach for automatic segmentation of cardiac structures in short-axis cine-MR images using deep neural networks [J ] . Expert Systems with Applications , 2022 , 197 : 116704 .
Painchaud N , Skandarani Y , Judge T , et al . Cardiac segmentation with strong anatomical guarantees [J ] . IEEE transactions on medical imaging , 2020 , 39 ( 11 ): 3703 - 3713 .
Silva I F S , Silva A C , Paiva A C , et al . A Multi-Stage Automatic Method Based on a Combination of Fully Convolutional Networks for Cardiac Segmentation in Short-Axis MRI [J ] . Applied Sciences (2076-3417), 2024 , 14 ( 16 ).
Fan J , Pei J , Bi X , et al . Context correlation aware network for cardiac segmentation [C ] // 2022 IEEE International Conference on Multimedia and Expo(ICME) . IEEE , 2022 : 1 - 6 .
Liu C , Xu J , Xiao Z , et al . An automatic cardiac magnetic resonance image segmentation algorithm based on deep learning [J ] . Computer Engineering & Science , 2022 , 44 ( 9 ): 1646 .
Ammar A , Bouattane O , Youssfi M . Automatic cardiac cine MRI segmentation and heart disease classification [J ] . Computerized Medical Imaging and Graphics , 2021 , 88 : 101864 .
Singh K R , Sharma A , Singh G K . MADRU-Net:Multiscale attention-based cardiac MRI segmentation using deep residual U-Net [J ] . IEEE Transactions on Instrumentation and Measurement , 2023 , 73 : 1 - 13 .
0
浏览量
0
下载量
0
CSCD
关联资源
相关文章
相关作者
相关机构
京公网安备11010802024621