MGM-3DUNet：多尺度边缘语义引导的图卷积序列脑肿瘤分割方法

庄建军; 李想; 景生华; 吕政隆

doi:10.11999/JEIT260128

MGM-3DUNet：多尺度边缘语义引导的图卷积序列脑肿瘤分割方法

doi: 10.11999/JEIT260128 cstr: 32379.14.JEIT260128

庄建军^{1, 2, ,},
李想¹,
景生华³,
吕政隆¹

1.
安徽建筑大学电子与信息工程学院合肥 230031
2.
南京信息工程大学电子与信息工程学院南京 210044
3.
东部战区总医院放疗科南京 210002

基金项目: 国家重点研发计划(2021YFE0105500)，国家自然科学基金(62171228)，“江苏高校‘青蓝工程’资助”

详细信息

作者简介:
庄建军：男，博士，教授，研究方向为深度学习、医疗人工智能

李想：男，硕士研究生，研究方向为深度学习、医学图像处理

景生华：女，硕士，副主任技师，研究方向为医学物理

吕政隆：男，硕士研究生，研究方向为医学图像处理、嵌入式

通讯作者:
庄建军　jjzhuang@nuist.edu.cn

中图分类号: TP391.41
计量
- 文章访问数: 4
- HTML全文浏览量: 3
- PDF下载量: 1
- 被引次数: 0
出版历程
- 修回日期: 2026-05-29
- 录用日期: 2026-05-29
- 网络出版日期: 2026-06-10

MGM-3DUNet: A Multi-scale Edge Semantic Guided Graph Convolutional Sequence Method for Brain Tumor Segmentation

1.
School of Electronics and Information Engineering, Anhui Jianzhu University, Hefei 230031, China
2.
School of Electronics and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
3.
Department of Radiation Oncology, Jinling Hospital, School of Medicine Nanjing University, Nanjing 210002, China

Funds: National Key Research and Development Program of China(2021YFE0105500), National Natural Science Foundation of China(62171228), “Jiangsu Provincial University ‘Blue and Yellow Project’ Funding”

摘要

摘要: 针对脑肿瘤尺度差异大、边界模糊、小病灶易漏检以及现有分割模型难以兼顾分割性能和参数规模等问题，该文以3DUNet为基线，提出了多尺度边缘语义引导的图卷积序列脑肿瘤分割方法——MGM-3DUNet。通过多尺度边缘语义引导模块(Multi-scale Edge Semantic Guidance Module, MEGM)将可学习的边缘检测与多尺度语义特征融合，精准捕捉肿瘤边界细节，生成边缘预测图提供辅助监督；通过图卷积序列模块(Graph Convolution Sequence Module, GCSM)融合图卷积的局部拓扑聚合能力与高效长程建模优势，缓解现有长序列特征建模方法在特征编码过程中局部细粒度细节易衰减、空间精细结构保留不足的问题；多尺度上下文感知模块(Multi-scale Context Perception Module，MCPM)对不同尺度特征动态加权融合，使解码器自适应匹配水肿区与核心区的特征需求，缓解尺度失衡导致的小病灶漏检。最后，在开源数据集BraTS2020、BraTS2021上进行了实验验证。结果表明，模型以2.3M的轻量级参数在整体肿瘤(WT)、肿瘤核心(TC)和增强肿瘤(ET)区域分别达到了91.2%、90.4%和89.2%的Dice系数，整体性能优于现有主流模型，在保持低计算成本的同时，实现了关键区域的精准分割。
- 3DUNet /
- 轻量级 /
- 边缘语义引导 /
- 图卷积 /
- 脑肿瘤分割
Abstract: Objective The model feature fusion method represented by U-Net and its three 3D variants is simplistic, and the segmentation of tumor core and enhancing tumor region is insufficiently fine-grained. Recent approaches such as VM-UNet have made progress in sequence modeling efficiency, but they focus more on global information modeling, and there are still deficiencies in local detail preservation and edge enhancement. Therefore, the current methods are still limited in segmentation accuracy and clinical utility. Methods MEGM is designed to enhance the segmentation accuracy of the tumor boundary through learnable edge detection. GCSM, which combines the local aggregation ability of graph convolution with the efficient long-range modeling advantages of Mamba-like structure, enhances semantic consistency while reducing parameters, and retains small tumor structure details. MCPM is introduced to improve the complementarity of tumor features at different scales through dual-scale fusion. Results and Discussions Experiments show that the average Dice and HD95 distances of the proposed method are better than those of the comparison method. The visualization results ( Figure 9, Figure 10) qualitatively confirm that the segmentation results are more accurate after incorporating MEGM. In summary, the method proposed in this paper demonstrates enhanced sensitivity to edge details and context correlation while maintaining low parameter count, and its segmentation performance is highly robust and accurate. Conclusions This method improves the accuracy of tumor boundary prediction by introducing edge enhancement in the shallow layer to emphasize tumor contours. In the bottleneck layer, multimodal local and global semantic information is fused, while multi-scale context features are integrated during the decoding stage. This design achieves high segmentation accuracy at low computational cost and is suitable for platform deployment with low computing power.
- 3DUNet /
- Lightweight /
- Edge semantic guidance /
- Graph convolution /
- Brain tumor segmentation

HTML全文

图 1 MGM-3DUNet网络结构

下载: 全尺寸图片幻灯片

图 2 MEGM结构

下载: 全尺寸图片幻灯片

图 3 GCSM结构

下载: 全尺寸图片幻灯片

图 4 MCPM结构

下载: 全尺寸图片幻灯片

图 5 不同尺度组合的整体及TC的Dice值

下载: 全尺寸图片幻灯片

图 6 不同尺度组合的计算效率

下载: 全尺寸图片幻灯片

图 7 4种模态示例

下载: 全尺寸图片幻灯片

图 8 数据集预处理流程图

下载: 全尺寸图片幻灯片

图 9 不同模块分割结果对比

下载: 全尺寸图片幻灯片

图 10 TC/WT边界对比

下载: 全尺寸图片幻灯片

图 11 分割结果可视化

下载: 全尺寸图片幻灯片

表 1 实验设备配置表

参数	配置
CPU	Intel(R)Core(TM)i7-14700HX(2.1GHz)
GPU	NVIDIA GeForce RTX 4060(8GB)
Windows	11
CUDA	11.8
PyTorch	2.5
Python	3.9

下载: 导出CSV

表 2 数据集划分及样本量分布

数据集名称	数据子集	样本数量	用途
BraTS2021	训练集	875	模型训练
	验证集	250	性能验证
	测试集	125	性能测试
BraTS2020	测试集	50	泛化验证

下载: 导出CSV

表 3 在BraTS2021数据集上使用基线模型3DUNet不同$ \alpha $时的结果对比

$ \alpha $	Dice/%				Sensitivity/%				HD95/mm
$ \alpha $	WT	TC	ET	均值	WT	TC	ET	均值	WT	TC	ET	均值
0.1	83.3	75.4	74.2	77.6	88.3	78.2	73.2	79.9	18.6	12.1	10.1	13.6
0.2	82.1	75.6	75.3	77.7	87.5	78.1	75.3	80.3	17.5	11.5	10.3	13.1
0.3	84.8	76.3	75.6	78.9	89.7	80.6	77.0	82.4	16.9	10.9	9.9	12.6
0.4	81.2	76.4	72.1	76.6	86.7	79.9	73.1	79.9	18.3	10.9	10.3	13.2
0.5	80.5	75.8	73.2	76.5	86.3	78.2	72.8	79.1	18.2	12.3	10.8	13.8
0.6	82.5	74.7	74.4	77.2	87.9	77.8	75.9	80.5	17.8	12.5	10.5	13.6
0.7	83.6	74.2	74.1	77.3	88.5	77.2	75.2	80.3	17.1	12.4	10.4	13.3
0.8	80.2	75.5	73.9	76.5	85.4	77.3	73.0	78.6	18.8	11.8	10.7	13.8
0.9	79.8	73.1	73.6	75.5	84.9	75.4	74.6	78.3	19.0	12.7	11.0	14.2
注：表中粗体表示最优值。

下载: 导出CSV

表 4 交叉验证结果对比

评价指标	Dice/%				Sensitivity/%				HD95/mm
评价指标	WT	TC	ET	均值	WT	TC	ET	均值	WT	TC	ET	均值
1	93.1	90.4	90.2	91.2	96.2	96.1	89.9	94.1	2.8	2.1	3.0	2.6
2	92.1	88.9	88.6	89.9	95.5	92.4	88.3	92.1	3.2	2.7	3.3	3.1
3	93.8	93.3	89.3	92.1	96.7	95.6	89.0	93.8	3.0	2.9	2.9	2.9
4	91.2	90.2	89.4	90.3	93.8	92.9	89.1	91.9	3.1	2.9	3.7	3.2
5	92.5	89.8	88.2	90.2	95.3	92.2	88.3	91.9	2.9	2.3	3.4	2.9
均值	92.5	90.5	89.1	90.7	95.5	93.8	88.9	92.7	3.0	2.6	3.3	3.0
注：表中所有数值均为平均值。

下载: 导出CSV

表 5 消融实验结果对比

模块	Dice/%				Sensitivity/%				HD95/mm
模块	WT	TC	ET	均值	WT	TC	ET	均值	WT	TC	ET	均值
3DUNet(1)	82.3	76.1	75.3	77.9	88.1	81.2	75.8	81.7	17.7	11.4	9.5	12.9
3DUNet+MEGM(2)	89.8	87.2	86.8	87.9	94.8	88.6	82.2	88.5	7.2	5.5	5.3	6.0
3DUNet+GCSM(3)	88.4	85.0	85.8	86.4	94.3	87.8	84.4	88.8	8.4	7.2	5.1	6.9
3DUNet+MCPM(4)	86.2	83.1	80.5	83.3	92.8	85.5	79.4	85.9	11.7	9.2	8.3	9.7
3DUNet+MEGM+GCSM(5)	90.2	88.8	88.2	89.1	95.3	90.2	89.8	91.8	6.4	4.5	4.2	5.0
3DUNet+MEGM+MCPM(6)	89.2	87.4	87.6	88.1	95.0	88.9	87.1	90.3	5.9	4.3	3.9	4.7
3DUNet+GCSM+MCPM(7)	88.9	88.2	86.2	87.8	94.2	88.2	88.6	90.3	6.2	5.1	4.5	5.3
MGM-3DUNet(8)	91.2	90.4	89.2	90.3	96.4	93.2	90.2	93.3	3.1	3.9	2.8	3.3
注：表中粗体表示最优值。

下载: 导出CSV

表 6 MGM-3DUNet与不同模型的对比结果

网络类型	模型	Dice/%				HD95/%				Param/M
网络类型	模型	WT	TC	ET	均值	WT	TC	ET	均值	Param/M
经典网络	UNet++^[21]	86.6	77.4	74.5	79.5	8.9	23.3	27.9	20.0	22.2
	V-Net^[6]	88.2	83.6	80.6	84.1	10.5	13.3	24.5	16.1	24.3
	MGM-3DUNet	91.2	90.4	89.2	90.3	3.1	3.9	2.8	3.3	2.3
目前主流网络	TransBTS^[9]	86.4	88.3	87.6	87.4	12.5	16.8	13.1	14.1	30.6
	UNETR^[22]	90.8	89.3	88.9	89.7	5.5	5.0	3.6	4.7	148.5
	Swin UNETR^[10]	90.8	89.5	89.0	89.8	5.4	4.8	3.7	4.6	35.5
	VM-UNet^[14]	90.2	85.6	80.2	85.3	5.4	7.8	5.3	6.2	18.7
	FS Inv-ResU-Net^[23]	90.5	86.5	82.8	86.6	5.5	4.5	2.4	4.1	26.3
	MR-SC-UNet^[24]	91.1	87.5	87.8	88.8	4.4	5.3	2.5	4.1	32.4
	DC-Seg^[27]	89.8	88.2	87.9	88.6	5.8	6.2	5.2	5.7	15.2
	VSMU-Net^[28]	90.4	89.3	88.6	89.4	6.1	4.8	5.0	5.3	25.1
	S²CA-Net^[29]	89.9	89.2	88.5	89.2	6.3	5.1	4.9	5.4	32.3
	MGM-3DUNet	91.2	90.4	89.2	90.3	3.1	3.9	2.8	3.3	2.3
轻量级网络	LATUP-Net^[25]	90.5	88.5	86.4	88.5	6.3	8.2	4.5	6.3	3.1
	ADHDC-Net^[26]	80.5	87.0	85.4	84.3	14.7	12.1	6.8	11.2	0.3
	MGM-3DUNet	91.2	90.4	89.2	90.3	3.1	3.9	2.8	3.3	2.3
注：表中粗体表示最优值

下载: 导出CSV

表 7 MGM-3DUNet在BraTS2020数据集上的对比结果

模型	Dice/%				HD95/%
模型	WT	TC	ET	均值	WT	TC	ET	均值
VM-UNet	87.2	85.9	82.4	85.2	6.2	8.3	6.4	7.0
LATUP-Net	87.5	88.1	85.3	87.0	5.5	9.2	5.3	6.7
TransBTS	82.5	83.6	81.8	82.6	15.5	20.3	6.8	14.2
DC-Seg	89.2	88.1	86.7	88.0	7.3	6.1	5.5	6.3
VSMU-Net	89.5	88.3	87.1	88.3	7.1	5.7	5.5	6.1
S²CA-Net	89.1	88.1	87.2	88.1	7.2	6.3	5.4	6.3
MGM-3DUNet	90.1	89.8	87.2	89.0	5.2	4.8	4.0	4.7
注：表中粗体表示最优值

下载: 导出CSV

参考文献(29)

[1]	KRISHNAPRIYA S and KARUNA Y. A survey of deep learning for MRI brain tumor segmentation methods: Trends, challenges, and future directions[J]. Health and Technology, 2023, 13(2): 181–201. doi: 10.1007/s12553-023-00737-3.
[2]	李锵, 阮方号, 关欣. 基于双路径特征融合的轻量级脑肿瘤分割网络[J]. 天津大学学报: 自然科学与工程技术版, 2024, 57(11): 1177–1186. doi: 10.11784/tdxbz202401015. LI Qiang, RUAN Fanghao, and GUAN Xin. Lightweight brain tumor segmentation network based on dual-path feature fusion[J]. Journal of Tianjin University: Science and Technology, 2024, 57(11): 1177–1186. doi: 10.11784/tdxbz202401015.
[3]	王晨, 杜晨曦, 刘瑞军, 等. 基于RGB-D图像的语义分割方法综述[J]. 计算机辅助设计与图形学学报, 2025, 37(1): 100–119. doi: 10.3724/SP.J.1089.2024-00107. WANG Chen, DU Chenxi, LIU Ruijun, et al. A review of semantic segmentation methods based on RGB-D images[J]. Journal of Computer-Aided Design & Computer Graphics, 2025, 37(1): 100–119. doi: 10.3724/SP.J.1089.2024-00107.
[4]	RONNEBERGER O, FISCHER P, and BROX T. U-Net: Convolutional networks for biomedical image segmentation[C]. Proceedings of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 2015: 234–241. doi: 10.1007/978-3-319-24574-4_28.
[5]	ÇIÇEK Ö, ABDULKADIR A, LIENKAMP S S, et al. 3D U-Net: Learning dense volumetric segmentation from sparse annotation[C]. Proceedings of the 19th International Conference on Medical Image Computing and Computer-Assisted Intervention, Athens, Greece, 2016: 424–432. doi: 10.1007/978-3-319-46723-8_49.
[6]	MILLETARI F, NAVAB N, and AHMADI S A. V-Net: Fully convolutional neural networks for volumetric medical image segmentation[C]. Proceedings of the 2016 Fourth International Conference on 3D Vision, Stanford, USA, 2016: 565–571. doi: 10.1109/3DV.2016.79.
[7]	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[C]. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 5998–6008.
[8]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[C]. Proceedings of the 9th International Conference on Learning Representations, 2021. (查阅网上资料, 未找到对应的出版地信息, 请确认).
[9]	WANG Wenxuan, CHEN Chen, DING Meng, et al. TransBTS: Multimodal brain tumor segmentation using Transformer[C]. Proceedings of the 24th International Conference on Medical Image Computing and Computer-Assisted Intervention, Strasbourg, France, 2021: 109–119. doi: 10.1007/978-3-030-87193-2_11.
[10]	HATAMIZADEH A, NATH V, TANG Yucheng, et al. Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images[C]. Proceedings of the 7th International Workshop Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, 2022: 272–284. doi: 10.1007/978-3-031-08999-2_22. (查阅网上资料,未找到对应的出版地信息,请确认).
[11]	HE Kelei, GAN Chen, LI Zhuoyuan, et al. Transformers in medical image analysis[J]. Intelligent Medicine, 2023, 3(1): 59–78. doi: 10.1016/j.imed.2022.07.002.
[12]	陈雷, 杨吉斌, 曹铁勇, 等. 一种基于Transformer特征金字塔的自蒸馏目标分割方法[J]. 电子与信息学报, 2025, 47(2): 551–560. doi: 10.11999/JEIT240735. CHEN Lei, YANG Jibin, CAO Tieyong, et al. A self-distillation object segmentation method based on Transformer feature pyramid[J]. Journal of Electronics & Information Technology, 2025, 47(2): 551–560. doi: 10.11999/JEIT240735.
[13]	MA Jun, LI Feifei, and WANG Bo. U-Mamba: Enhancing long-range dependency for biomedical image segmentation[EB/OL]. https://arxiv.org/abs/2401.04722, 2024.
[14]	RUAN Jiacheng, LI Jincheng, XIANG Suncheng, et al. VM-UNet: Vision mamba UNet for medical image segmentation[J]. ACM Transactions on Multimedia Computing, Communications and Applications, 2025. doi: 10.1145/3767748.
[15]	丁建睿, 张听, 刘家栋, 等. 融合邻域注意力和状态空间模型的医学视频分割算法[J]. 电子与信息学报, 2025, 47(5): 1582–1595. doi: 10.11999/JEIT240755. DING Jianrui, ZHANG Ting, LIU Jiadong, et al. A medical video segmentation algorithm integrating neighborhood attention and state space model[J]. Journal of Electronics & Information Technology, 2025, 47(5): 1582–1595. doi: 10.11999/JEIT240755.
[16]	孙军梅, 葛青青, 李秀梅, 等. 一种具有边缘增强特点的医学图像分割网络[J]. 电子与信息学报, 2022, 44(5): 1643–1652. doi: 10.11999/JEIT210784. SUN Junmei, GE Qingqing, LI Xiumei, et al. A medical image segmentation network with boundary enhancement[J]. Journal of Electronics & Information Technology, 2022, 44(5): 1643–1652. doi: 10.11999/JEIT210784.
[17]	周涛, 党培, 陆惠玲, 等. 跨模态跨尺度跨维度的PET/CT图像的Transformer分割模型[J]. 电子与信息学报, 2023, 45(10): 3529–3537. doi: 10.11999/JEIT221204. ZHOU Tao, DANG Pei, LU Huiling, et al. A Transformer segmentation model for PET/CT images with cross-modal, cross-scale and cross-dimensional[J]. Journal of Electronics & Information Technology, 2023, 45(10): 3529–3537. doi: 10.11999/JEIT221204.
[18]	SUDRE C H, LI Wenqi, VERCAUTEREN T, et al. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations[C]. Proceedings of the 3rd International Workshop Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, Québec City, Canada, 2017: 240–248. doi: 10.1007/978-3-319-67558-9_28.
[19]	马金林, 杨继鹏. 基于TT分解的轻量化肝肿瘤分割方法[J]. 电子与信息学报, 2026, 48(1): 335–345. doi: 10.11999/JEIT250293. MA Jinlin and YANG Jipeng. Tensor-train decomposition for lightweight liver tumor segmentation[J]. Journal of Electronics & Information Technology, 2026, 48(1): 335–345. doi: 10.11999/JEIT250293.
[20]	周涛, 叶鑫宇, 刘凤珍, 等. 利用跨模态轻量级YOLOv5模型的PET/CT肺部肿瘤检测[J]. 电子与信息学报, 2024, 46(2): 624–632. doi: 10.11999/JEIT230052. ZHOU Tao, YE Xinyu, LIU Fengzhen, et al. CL-YOLOv5: PET/CT lung cancer detection with cross-modal lightweight YOLOv5 model[J]. Journal of Electronics & Information Technology, 2024, 46(2): 624–632. doi: 10.11999/JEIT230052.
[21]	ZHOU Zongwei, RAHMAN SIDDIQUEE M, TAJBAKHSH N, et al. UNet++: A nested U-Net architecture for medical image segmentation[C]. Proceedings of the 4th Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, Granada, Spain, 2018: 3–11. doi: 10.1007/978-3-030-00889-5_1.
[22]	HATAMIZADEH A, TANG Yucheng, NATH V, et al. UNETR: Transformers for 3D medical image segmentation[C]. Proceedings of the 2022 IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, USA, 2022: 1748–1758. doi: 10.1109/WACV51458.2022.00181.
[23]	阮东升, 施哲彬, 王嘉辉, 等. 结合跨模态特征激励与双分支交叉注意力融合的左心房疤痕分割方法[J]. 电子与信息学报, 2025, 47(5): 1596–1608. doi: 10.11999/JEIT240775. RUAN Dongsheng, SHI Zhebin, WANG Jiahui, et al. Left atrial scar segmentation method combining cross-modal feature excitation and dual branch cross attention fusion[J]. Journal of Electronics & Information Technology, 2025, 47(5): 1596–1608. doi: 10.11999/JEIT240775.
[24]	LIU Jiangjiang, HOU Qibin, LIU Zhiang, et al. PoolNet+: Exploring the potential of pooling for salient object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(1): 887–904. doi: 10.1109/TPAMI.2021.3140168.
[25]	ALWADEE E J, SUN Xianfang, QIN Yipeng, et al. LATUP-Net: A lightweight 3D attention U-Net with parallel convolutions for brain tumor segmentation[J]. Computers in Biology and Medicine, 2025, 184: 109353. doi: 10.1016/j.compbiomed.2024.109353.
[26]	LIU Hengxin, HUO Guoqiang, LI Qiang, et al. Multiscale lightweight 3D segmentation algorithm with attention mechanism: Brain tumor image segmentation[J]. Expert Systems with Applications, 2023, 214: 119166. doi: 10.1016/j.eswa.2022.119166.
[27]	LI Haitao, LI Ziyu, MAO Yiheng, et al. DC-Seg: Disentangled contrastive learning for brain tumor segmentation with missing modalities[C]. Proceedings of the 28th International Conference on Medical Image Computing and Computer Assisted Intervention, Daejeon, South Korea, 2026: 138–148. doi: 10.1007/978-3-032-04984-1_14.
[28]	ZHU Zhiqin, WANG Zhandong, QI Guanqiu, et al. Visually stabilized Mamba U-shaped network with strong inductive bias for 3-D brain tumor segmentation[J]. IEEE Transactions on Instrumentation and Measurement, 2025, 74: 2518511. doi: 10.1109/TIM.2025.3551581.
[29]	ZHOU Lifang, JIANG Yu, LI Weisheng, et al. Shape-scale co-awareness network for 3D brain tumor segmentation[J]. IEEE Transactions on Medical Imaging, 2024, 43(7): 2495–2508. doi: 10.1109/TMI.2024.3368531.