PAC-UNet：融合双路径卷积与注意力机制的脑肿瘤分割网络

PAC-UNet: a brain tumor segmentation network integrating dual-path convolution and attention mechanism

摘要

针对脑肿瘤MRI影像中病灶边界模糊、小目标分割敏感性不足及多尺度上下文信息捕获不充分等问题，提出一种融合双路径卷积模块与混合注意力机制模块的U型分割网络PAC-UNet。该网络采用新型双路径卷积模块替换标准卷积，通过并行结构整合动态核卷积与风车状卷积路径，其中，动态核卷积路径结合径向基函数RBF核生成器替代传统固定卷积核，通过可微分样条参数化实现核函数的动态适应，平衡局部细节特征与全局语义表征；风车状卷积路径采用四并行结构有效扩展感受野，增强上下文信息捕获能力。同时利用通道与空间双重注意力机制融合双路径特征，聚焦关键特征并抑制冗余背景干扰，协同优化多尺度特征表达。在BraTS2023数据集上的实验结果表明，PAC-UNet能够显著提升脑肿瘤分割性能。

Abstract

To address the issues of blurred lesion boundaries in brain tumor MRI images, insufficient segmentation sensitivity for small targets, and limited ability to extract multi-scale contextual information, this study proposes a U-shaped segmentation network named PAC-UNet, which integrates a dual-path convolution module and a hybrid attention mechanism module. The network uses a novel dual-path convolution module to replace standard convolutions, enabling parallel fusion of dynamic kernel convolution and pinwheel convolution paths. In the dynamic kernel convolution path, a radial basis function kernel generator replaces traditional fixed convolution kernels, which achieves dynamic kernel function adaptation through differentiable spline parameterization, thus balancing local detailed features and global semantic representations. The pinwheel convolution path adopts a four-parallel structure to effectively expand the receptive field and enhance the ability to capture contextual information. Meanwhile, a channel-spatial dual attention mechanism is used to fuse dual-path features, with a focus on key features while suppressing redundant background interference, thereby collaboratively optimizing multi-scale feature representations. Experiment results on the BraTS2023 dataset show that PAC-UNet achieves significant improvements in brain tumor segmentation performance.

关键词

脑肿瘤分割 / 双路径卷积 / 风车状卷积 / 动态核卷积 / 注意力机制 / U-Net

Key words

brain tumor segmentation / dual-path convolution / pinwheel convolution / dynamic kernel convolution / attention mechanism / U-Net

参考文献

[1] Wen PY, Packer RJ . The 2021 WHO classification of tumors of the central nervous system: clinical implications[J]. Neuro Oncol, 2021, 23(8): 1215-1217.

[2] Menze BH, Jakab A, Bauer S, et al. The multimodal brain tumor image segmentation benchmark (BRATS)[J]. IEEE Trans Med Imaging, 2015, 34(10): 1993-2024.

[3] Kickingereder P, Isensee F, Tursunova I, et al. Automated quantitative tumour response assessment of MRI in neuro—oncology with artificial neural networks: a multicentre, retrospective study[J]. Lancet Oncol, 2019, 20(5): 728-740.

[4] Ronneberger O, Fischer P, Brox T . U—Net: convolutional networks for biomedical image segmentation[C]// Medical Image Computing and Computer—Assisted Intervention—MICCAI 2015. Cham: Springer International Publishing, 2015: 234-241.

[5] Zhang JX, Jiang ZK, Dong J, et al. Attention gate ResU—Net for automatic MRI brain tumor segmentation[J]. IEEE Access, 2020, 8: 58533-58545.

[6] Jin L. 3AU—Net: triple attention U—Net for retinal vessel segmentation[C]// 2020 IEEE 2nd International Conference on Civil Aviation Safety and Information Technology (ICCASIT). Piscataway, NJ, USA: IEEE, 2020: 612-615.

[7] Gao CR, Cheng JL, Yang ZY, et al. SAA—NET: a medical image segmentation framework based on stream—across attention[C]// 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). Piscataway, NJ, USA: IEEE, 2023: 1-5.

[8] Hu JF, Wang H, Wang J, et al. SA—Net: a scale—attention network for medical image segmentation[J]. PLoS One, 2021, 16(4): e0247388.

[9] Gu ZW, Cheng J, Fu HZ, et al. CE—Net: context encoder network for 2D medical image segmentation[J]. IEEE Trans Med Imaging, 2019, 38(10): 2281-2292.

[10] Wang QL, Wu BG, Zhu PF, et al. ECA—Net: efficient channel attention for deep convolutional neural networks[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ, USA: IEEE, 2020: 11531-11539.

[11] Chen LP, Jia DX, Gao H, et al. SkaNet: split kernel attention network[C]// Artificial Neural Networks and Machine Learning—ICANN 2023. Cham: Springer Nature Switzerland, 2023: 459-473.

[12] Shen X, Han DZ, Chang CC, et al. GFSNet: Gaussian Fourier with sparse attention network for visual question answering[J]. Artif Intell Rev, 2025, 58(6): 159.

[13] Song H, Song YH, Zhang YL . SCA net: sparse channel attention module for action recognition[C]// 2020 25th International Conference on Pattern Recognition (ICPR). Piscataway, NJ, USA: IEEE, 2021: 1189-1196.

[14] Li CX, Liu XY, Li WY, et al. U—KAN makes strong backbone for medical image segmentation and generation[EB/OL]. (2024—08—22). https://arxiv.org/abs/2406.02918.

[15] Shen LF, Wang Q, Zhang YJ, et al. DSKCA—UNet: dynamic selective kernel channel attention for medical image segmentation[J]. Medicine (Baltimore), 2023, 102(39): e35328.

[16] Huang GC, Chen ZP, Chao Y, et al. ALU—Net: an attention—based mechanism lightweight U—Net for multimodal brain tumor segmentation[C]// Proceedings Volume 13105, International Conference on Computer Graphics, Artificial Intelligence, and Data Processing (ICCAID 2023). 2024: 131050M.

[17] Cheng AH . Radial basis function collocation method[C]// Computational Mechanics. Berlin, Heidelberg: Springer Berlin Heidelberg, 2009: 219.

[18] Fey M, Lenssen JE, Weichert F, et al. SplineCNN: fast geometric deep learning with continuous B—spline kernels[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, NJ, USA: IEEE, 2018: 869-877.

[19] Yang JN, Liu SL, Wu JJ, et al. Pinwheel—shaped convolution and scale—based dynamic loss for infrared small target detection[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2025, 39(9): 9202-9210.

[20] Mohan G, Subashini MM . MRI based medical image analysis: survey on brain tumor grade classification[J]. Biomed Signal Process Control, 2018, 39: 139-161.

[21] Milletari F, Navab N, Ahmadi SA . V—Net: fully convolutional neural networks for volumetric medical image segmentation[C]// 2016 Fourth International Conference on 3D Vision (3DV). Piscataway, NJ, USA: IEEE, 2016: 565-571.

[22] Isensee F, Jäger PF, Full PM, et al. nnU—Net for brain tumor segmentation[C]// Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. Cham: Springer International Publishing, 2021: 118-132.

[23] Bakas S. BraTS 2023 challenge protocol, synapse, 2023[EB/OL]. https://www.synapse.org/brats2023.

[24] Zhou ZW, Rahman Siddiquee MM, Tajbakhsh N, et al. UNet++: a nested U—Net architecture for medical image segmentation[C]// Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Cham: Springer International Publishing, 2018: 3-11.

[25] Oktay O, Schlemper J, Le Folgoc L, et al. Attention U—Net: learning where to look for the pancreas[EB/OL]. (2018—05—20). https://arxiv.org/abs/1804.03999.

[26] Chen JN, Lu YY, Yu QH, et al. TransUNet: transformers make strong encoders for medical image segmentation[EB/OL]. (2021—02—08). https://arxiv.org/abs/2102.04306.

[27] Cao H, Wang YY, Chen J, et al. Swin—Unet: Unet—like pure transformer for medical image segmentation[C]// Computer Vision—ECCV 2022 Workshops. Cham: Springer Nature Switzerland, 2023: 205-218.

[28] Zhou RJ, Qu LS, Zhang L, et al. Fed—MUnet: multi—modal federated Unet for brain tumor segmentation[C]// 2024 IEEE International Conference on E—health Networking, Application & Services (HealthCom). Piscataway, NJ, USA: IEEE, 2024: 1-6.

[29] Yang LJ, Dong QM, Lin D, et al. MUNet: a novel framework for accurate brain tumor segmentation combining UNet and mamba networks[J]. Front Comput Neurosci, 2025, 19: 1513059.

基金资助

国家自然科学基金(61601173)

河南理工大学博士创新基金(760807/013)