山东中医药大学医学信息工程学院(济南 250355)
魏国辉。E-mail:bmie530@163.com
收稿:2026-01-05,
修回:2026-03-25,
纸质出版:2026-06-25
移动端阅览
闫家奕,魏国辉.基于多尺度特征提取与多特征融合的甲状腺结节超声影像分割[J].北京生物医学工程,2026,45(3):229-238.
YAN Jiayi,WEI Guohui.Ultrasound image segmenation of thyroid nodules based on multi-scale feature extraction and multi-feature fusion[J].Beijing Biomedical Engineering,2026,45(3):229-238.
闫家奕,魏国辉.基于多尺度特征提取与多特征融合的甲状腺结节超声影像分割[J].北京生物医学工程,2026,45(3):229-238. DOI: 10.3969/j.issn.1002-3208.2026.03.002.
YAN Jiayi,WEI Guohui.Ultrasound image segmenation of thyroid nodules based on multi-scale feature extraction and multi-feature fusion[J].Beijing Biomedical Engineering,2026,45(3):229-238. DOI: 10.3969/j.issn.1002-3208.2026.03.002.
目的
2
针对甲状腺结节超声影像分割中结节尺寸差异显著、图像低对比度及强斑点噪声干扰的问题,提出一种基于多尺度特征提取与全局多特征融合的精准分割方法,旨在提升不同尺度结节的识别能力与复杂背景下的泛化性能。
方法
2
采用经典的编码器-解码器网络架构作为基础框架。在此基础上,集成多尺度残差连接模块(multi-scale residual connection module ,MRCM),用于捕获传统固定感受野卷积核难以同时捕获到的多尺度结节特征。同时,引入改进的全局金字塔引导模块(improved global pyramid guidance ,IGPG),用于完成高效且去噪的全局多特征融合。使用本文所构建的方法在TN3K和DDTI两个公开数据集上进行了验证。
结果
2
MRCM通过密集堆叠3个级联的3×3卷积核,实现感受野的连续扩展,使其能够自适应地捕捉从精细局部纹理到更大范围上下文信息
的不同尺度特征,有效应对结节尺寸的剧烈变化。IGPG 摒弃空洞卷积,转而利用 MRCM 进行跨层特征融合,并以深层语义信息作为空间注意力向导,有效筛选并整合多阶段特征,从而显著抑制噪声干扰,增强判别性特征的表达能力。在TN3K数据集上,本文算法在交并比(intersection over union,IoU)、95%豪斯多夫距离(Hausdorff distance-95% ,HD95)和
F
1分数上分别达到了72.21%、18.49和80.97%,相较于U-Net基准,IoU提升了4.0%,
F
1提升了3.5%。在DDTI数据集上,本文算法的IoU、HD95和
F
1分数分别为63.40%、18.61和74.97%,相较于U-Net基准,IoU提升了3.0%,
F
1提升了2.4%。结果均显著优于FCN、U-Net、MultiResUNet、SGUNet和DC-UNet等现有主流方法,并在消融实验中证明了MRCM和 IGPG 模块的有效性和协同作用。
结论
2
基于MRCM和IGPG的分割方法通过协同作用,可有效解决甲状腺结节超声影像分割中的尺度变异和噪声干扰问题,显著提升分割精度和鲁棒性,为甲状腺结节计算机辅助诊断系统提供新的技术方案。
Objective
2
To address the challenges of significant variations in nodule size, low image contrast, and strong speckle noise in ultrasound image segmentation of thyroid nodules, this paper proposes a precise segmentation method based on multi-scale feature extraction and global multi-feature fusion, aiming to enhance the recognition capability of features at different scales and improve generalization performance under complex backgrounds.
Methods
2
A classical encoder-decoder network architecture is adopted as the foundational framework. Building upon this, a Multi-scale Residual Connection Module (MRCM) is integrated to capture multi-scale nodule features that traditional fixed-receptive-field convolution kernels struggle to capture simultaneously. Concurrently, an Improved Global Pyramid Guidance (IGPG) module is introduced to achieve efficient and denoised global multi-feature fusion. The developed method was validated on two public datasets, TN3K and DDTI.
Results
2
The MRCM achieves continuous expansion of the receptive field by densely stacking three cascaded 3×3 convolution kernels, enabling it to adaptively capture features at different scales—from fine local textures to broader contextual information—effectively addressing drastic variations i
n nodule size. The IGPG module abandons dilated convolutions, instead utilizing MRCM for cross-layer feature fusion, and employs deep semantic information as a spatial attention guide to effectively filter and integrate multi-stage features, thereby significantly suppressing noise interference and enhancing the expression ability of discriminative features. On the TN3K dataset, the algorithm achieved IoU, HD95, and
F
1 scores of 72.21%, 18.49, and 80.97% respectively. Compared to the U-Net baseline, IoU increased by 4.0% and
F
1 by 3.5%. On the DDTI dataset, the algorithm's IoU, HD95, and
F
1 scores were 63.40%, 18.61, and 74.97% respectively. Compared to the U-Net baseline, IoU increased by 3.0% and
F
1 by 2.4%. The results significantly outperformed existing mainstream methods such as FCN, U-Net, MultiResUNet, SGUNet, and DC-UNet, and the effectiveness and synergistic effect of the MRCM and IGPG modules were demonstrated in ablation experiments.
Conclusions
2
By synergistically integrating the MRCM and IGPG modules, the proposed segmentation method effectively addresses the issues of scale variation and noise interference in ultrasound image segmentation of thyroid nodules, markedly improving segmentation accuracy and robustness, thereby offering a novel technical solution for computer-aided diagnosis systems of thyroid nodules.
中华医学会内分泌学分会,中华医学会外科学分会甲状腺及代谢外科学组,中国抗癌协会头颈肿瘤专业委员会,等.甲状腺结节和分化型甲状腺癌诊治指南(第二版)[J].中华内分泌代谢杂志,2023,39(3):181-226.
中华医学会超声医学分会浅表器官和血管学组,中国甲状腺与乳腺超声人工智能联盟.2020甲状腺结节超声恶性危险分层中国指南:C-TIRADS[J].中华超声影像学杂志,2021,30(3):185-200.
Chen WQ,Zheng RS,Zhang SW,et al.Cancer incidence and mortality in China in 2013:an analysis based on urbanization level[J].Chinese Journal of Cancer Research,2017,29(1):1-10.
Chen JY,You HJ,Li K.A review of thyroid gland segmentation and thyroid nodule segmentation methods for medical ultrasound images[J].Computer Methods and Programs in Biomedicine, 2020,185:105329.
Tao Z,Dang H,Shi YT,et al.Local and context-attention adaptive LCA-Net for thyroid nodule segmentation in ultrasound images[J].Sensors,2022,22(16):5984.
Durante C,Grani G,Lamartina L,et al.The diagnosis and management of thyroid nodules:a review[J].JAMA,2018,319(9):914-924.
Chi JN,Li ZL,Sun ZY,et al.Hybrid transformer UNet for thyroid segmentation from ultrasound scans[J].Computers in Biology and Medicine,2022,153:106453.
Song PF,Li JJ,Fan H,et al.TGDAUNet:Transformer and GCNN based dual-branch attention UNet for medical image segmentation[J].Computers in Biology and Medicine,2023,167:107583.
Lafraxo S,El Ansari M,Koutti L,et al.SEDARU-net:a squeeze-excitation dilated based residual U-Net with attention mechanism for automatic melanoma lesion segmentation[J].Multimedia Tools and Applications,2025,84(21):23935-23962.
Haribabu K,Prasath R,Ir P.MLRT-UNet:An Efficient Multi-Level Relation Transformer Based U-Net for Thyroid Nodule Segmentation[J].Computer Modeling in Engineering & Sciences, 2025,143(1):413-448.
Liu Z,Lin YT,Cao Y,et al.Swin transformer:Hierarchical vision transformer using shifted windows[C]//IEEE/CVF (Eds.). Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE, 2021: 10012-10022.
Vaswani A,Shazeer N,Parmar N,et al.Attention is all you need[C]// 31st Conference on Neural Information Processing Systems (NIPS 2017). Long Beach, CA, USA: Curran Associates, 2017: 6000-6010.
Dosovitskiy A,Beyer L,Kolesnikov A,et al.An image is worth 16x16 words:Transformers for image recognition at scale[J]. arXiv: 2010.11929 .
曹玉珍,郑洁,余辉,等. 基于超声影像的甲状腺结节智能分割算法[J].天津大学学报,2022,55(7):674-681.
Cao YZ,Zheng J,Yu H,Wang F,Zhang J.Intelligent segmentation of thyroid nodules based on ultrasound images[J].Journal of Tianjin University(Science and Technology),2022,55(7):674-681.
Liu J,Peng YL,Pu MS,et al.KTUNet:A hybrid CNN-Transformer with kolmogorov-arnold convolution for thyroid nodule segmentation in ultrasound images[J].Digital Medicine, 2025,11(2):e24-00019.
Bi H,Cai CJ,Sun JW,et al.BPAT-UNet:Boundary preserving assembled transformer UNet for ultrasound thyroid nodule segmentation[J].Computer Methods and Programs in Biomedicine, 2023,238:107614.
Deng XY,Dang ZY,Pan LH.STU3Net:An improved U‐net with swin transformer fusion for thyroid nodule segmentation[J].International Journal of Imaging Systems and Technology, 2024,34(5):23160.
Ibtehaz N,Rahman MS.MultiResUNet:rethinking the U-Net architecture for multimodal biomedical image segmentation[J].Neural Networks,2020,121:74-87.
Zhu SQ,Yu CR,Hu JJ.Regularizing deep neural networks for medical image analysis with augmented batch normalization[J].Applied Soft Computing,2024,154:111337.
Wang HC,Fu TF,Du YQ,et al.Scientific discovery in the age of artificial intelligence[J]. Nature,2023,620:47-60.
Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the inception architecture for computer vision[C]//IEEE eds.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, USA: IEEE, 2016: 2818-2826.
Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning[C]//AAAI eds.Proceedings of the AAAI Conference on Artificial Intelligence. San Francisco, CA, USA: AAAI Press, 2017.
Feng SL,Zhao HM,Shi F,et al.CPFNet: Context pyramid fusion network for medical image segmentation[J].IEEE Transactions on Medical Imaging,2020,39(10):3008-3018.
Chen XN,Liang C,Huang D,et al.Symbolic discovery of optimization algorithms[J].Advances in Neural Information Processing Systems,2023,36:49205-49233.
Valanarasu JMJ, Patel VM. Unext: Mlp-based rapid medical image segmentation network[C]//MICCAI eds.International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature, 2022: 23-33.
Gong HF, Chen GQ, Wang RR, et al. Multi-task learning for thyroid nodule segmentation with thyroid region prior[C]//IEEE eds. 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). Nice, France: IEEE, 2021: 257-261.
Pedraza L, Vargas C, Narváez F, et al. An open access thyroid ultrasound image database[C]//SPIE eds. 10th International Symposium on Medical Information Processing and Analysis. Bogota, Colombia: SPIE, 2015: 188-193.
Pan HT, Zhou Q, Latecki LJ. Sgunet: Semantic guided unet for thyroid nodule segmentation[C]//IEEE eds.2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). Nice, France: IEEE, 2021: 630-634.
Lou AG, Guan SY, Loew M. DC-UNet: rethinking the U-Net architecture with dual channel efficient CNN for medical image segmentation[C]//SPIE eds. Medical Imaging 2021: Image Processing San Diego, CA, USA: SPIE, 2021: 758-768.
Cao H, Wang YY, Chen J, et al. Swin-unet: Unet-like pure transformer for medical image segmentation[C]//ECCV eds.European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 205-218.
0
浏览量
18
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