基于Mamba结构的轻量级皮肤病变图像分割网络

doi:10.11996/JG.j.2095-302X.2025061257

摘要/Abstract

摘要：

皮肤病变分割是医学图像分析中的一项重要任务，对于皮肤类疾病的早期诊断和治疗具有重要意义。然而，在处理高分辨率皮肤图像和捕捉细微病变特征时，现有模型仍面临着计算复杂度高以及冗余信息处理不足等挑战。为此，提出一种基于Mamba结构的轻量级皮肤病变图像分割网络ResMamba，采用六级U型结构，主要通过将Mamba嵌入到视觉状态空间中同时引入到编解码器中，ResVSS模块作为编码器的核心组成部分，通过删除冗余线性层可减少参数量，同时结合深度卷积块和可学习尺度参数对残差连接进行缩放，从而通过降低模型复杂度来提升分割精度。在跳跃连接模块使用多级、多尺度信息融合模块生成空间和通道注意力图，有效融合了多尺度信息。通过在公开皮肤数据集ISIC2017和ISIC2018上进行实验验证，结果表明，ResMamba模型在平衡参数数量和分割性能方面都具有较好的分割性能，验证了该模型的有效性。

关键词: 深度学习, 皮肤病变分割, Mamba结构, 状态空间模型, 轻量化

Abstract:

Segmentation of skin lesions is an important task in medical image analysis, and is of great significance for the early diagnosis and treatment of skin diseases. However, when processing high-resolution skin images and capturing subtle lesion features, existing models still face challenges such as high computational complexity and insufficient processing of redundant information. To address this end, a lightweight skin lesion image segmentation network based on the Mamba structure was proposed, ResMamba adopted a six-level U-shaped structure, embedding Mamba into the visual state space and introducing it into the codec. The ResVSS module, as the core component of the encoder, reduced the number of parameters by removing a redundant linear layer, and at the same time combined the deep convolution block and learnable scale parameters to scale the residual connection, thereby reducing the complexity of the model while improving the segmentation accuracy. In the hopping connection module, a multi-level multi-scale information fusion module was used to generate spatial and channel attention maps, which effectively fused multi-scale information. Through experimental verification on the public skin dataset ISIC2017 and ISIC2018, the results demonstrated that the ResMamba model achieved good segmentation performance in terms of the number of balance parameters and segmentation performance, thus verifying the effectiveness of the model.

Key words: deep learning, skin lesion segmentation, mamba structure, state space models, lightweight

中图分类号:

贺蒙蒙, 张小艳, 李洪安. 基于Mamba结构的轻量级皮肤病变图像分割网络[J]. 图学学报, 2025, 46(6): 1257-1266.

HE Mengmeng, ZHANG Xiaoyan, LI Hongan. Lightweight skin lesion image segmentation network based on Mamba structure[J]. Journal of Graphics, 2025, 46(6): 1257-1266.

图/表 11

图1 ResMamba网络模型结构图((a) 本文模型；(b) 跳跃连接路径的多阶段和多尺度信息融合模块结构)

Fig. 1 Structure diagram of the ResMamba network model ((a) The model in this paper; (b) Multi-stage and multi-scale information fusion module structure of skip connection paths)

图2 RM Layer架构

Fig. 2 RM Layer Architecture

图3 ResVSS和VSS模块

Fig. 3 ResVSS and VSS modules ((a) ResVSS; (b) VSS)

表1 ISIC2017和ISIC2018数据集上训练集、验证集和测试集分割图像数据

Table 1 Training set, validation set and test set segmented image data on ISIC2017 and ISIC2018 datasets

Dateset	Train	Valid	Test	Total
ISIC2017	2 000	150	600	2 750
ISIC2018	1 815	259	520	2 594

表2 不同方法在ISIC2017数据集上的实验结果

Table 2 Experimental results of different methods on ISIC2017 dataset

Model	Params/M	GFLOPs	mIoU↑	DSC↑	ACC↑	Sen↑	Spe↑
U-Net	7.770	13.780	79.29	83.95	95.65	86.82	97.28
TransFuse	26.270	11.530	79.21	88.40	96.17	87.14	97.98
MALUNet	0.175	0.083	78.78	88.13	96.18	84.78	98.47
VM-UNet	27.430	4.112	80.23	89.03	96.62	86.89	97.45
VM-UNetV2	12.380	2.473	81.34	89.73	96.85	87.39	97.38
LightM-UNet	1.270	0.267	82.18	90.22	96.70	88.73	98.67
ResMamba (Ours)	0.043	0.059	83.89	89.95	96.99	89.21	98.73

表3 不同方法在ISIC2018数据集上的实验结果

Table 3 Experimental results of different methods on ISIC2018 dataset

Model	Params/M	GFLOPs	mIoU↑	DSC↑	ACC↑	Sen↑	Spe↑
U-Net	7.770	13.780	79.56	84.75	94.05	85.86	96.69
TransFuse	26.270	11.530	80.63	89.27	95.66	89.28	95.74
MALUNet	0.175	0.083	80.25	89.04	95.62	89.64	96.19
VM-UNet	27.430	4.112	81.35	89.71	96.19	88.49	96.93
VM-UNetV2	12.380	2.473	81.37	89.73	96.37	87.75	97.61
LightM-UNet	1.270	0.267	82.71	89.32	96.83	89.69	97.89
ResMamba (Ours)	0.043	0.059	84.37	90.21	97.04	89.62	97.92

图4 ISIC2017 (前3行)和ISIC2018 (后3行)数据集分割可视化展示

Fig. 4 ISIC2017 (first three rows) and ISIC2018 (last three rows) dataset segmentation visualizations are shown ((a) Original image; (b) Ground Truth; (c) U-Net; (d) Transfuse; (e) MALUNet; (f) VM-UNet; (g) VM-UNetV2; (h) LightM-UNet; (i) ResMamba (Ours))

图5 ISIC2017数据集Pramas与各模型mIoU结果对比

Fig. 5 Comparison of ISIC2017 dataset Pramas with mIoU results for each model

图6 ISIC2017数据集GFLOPs与各模型mIoU结果对比

Fig. 6 Comparison of ISIC2017 dataset GFLOPs with mIoU results for each model

图7 ISIC2017数据集消融实验分割可视化展示

Fig. 7 ISIC2017 dataset ablation study segmentation visualizations are shown ((a) Original image; (b) Ground Truth; (c) U-Net; (d) U-Net+SAB+CAB; (e) U-Net+RM Layer; (f) ResMamba (Ours))

表4 消融实验结果

Table 4 Results of ablation experiments

Model	Params/M	GFLOPs	mIoU↑	DSC↑	ACC↑	Sen↑	Spe↑
U-Net	7.770	13.780	79.29	83.95	95.66	86.82	97.28
U-Net+SAB+CAB	10.890	6.365	80.32	88.74	96.01	87.83	97.96
U-Net+RM Layer	0.127	0.074	81.49	89.06	96.04	89.05	98.34
ResMamba (Ours)	0.043	0.059	83.89	89.95	96.99	89.21	98.73

参考文献 23

[1]	文思佳, 张栋, 赵伟强, 等. 融合CNN-Transformer的医学图像分割网络[J]. 计算机与数字工程, 2024, 52(8): 2452-2456.
	WEN S J, ZHANG D, ZHAO W Q, et al. Medical image segmentation network integrated with CNN-Transformer[J]. Computer and Digital Engineering, 2024, 52(8): 2452-2456 (in Chinese).
[2]	FAN Y Z, SONG J H, YUAN L, et al. HCT-Unet: multi-target medical image segmentation via a hybrid CNN-Transformer Unet incorporating multi-axis gated multi-layer perceptron[J]. The Visual Computer, 2024, 41(5): 3457-3472. DOI
[3]	熊岚堃, 张桂梅, 刘晖群, 等. 结合轴向增强Transformer与CNN双编码的医学图像分割方法[EB/OL]. (2025-02-06) [2025-02-08]. http://kns.cnki.net/kcms/detail/11.2925.TP.2025 0206.1623.033.html.
	XIONG L K, ZHANG G M, LIU H Q, et al. Combination of axial enhanced Transformer and CNN network for medical image segmentation[EB/OL]. (2025-02-06) [2025-02-08]. http://kns.cnki.net/kcms/detail/11.2925.TP.2025 0206.1623.033.html. (in Chinese).
[4]	RONNEBERGER O, FISCHER P, BROX T. U-net: convolutional networks for biomedical image segmentation[C]// The 18th International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer, 2015: 234-241.
[5]	李翠云, 白静, 郑凉. 融合边缘增强注意力机制和U-Net网络的医学图像分割[J]. 图学学报, 2022, 43(2): 273-278.
	LI C Y, BAI J, ZHENG L. A U-Net based contour enhanced attention for medical image segmentation[J]. Journal of Graphics, 2022, 43(2): 273-278 (in Chinese). DOI
[6]	CHEN J N, LU Y Y, YU Q H, et al. TransUNet: transformers make strong encoders for medical image segmentation[EB/OL]. (2021-02-08) [2024-08-04]. https://arxiv.org/abs/2102.04306.
[7]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: transformers for image recognition at scale[EB/OL]. (2021-06-03) [2024-08-04]. https://dblp.uni-trier.de/db/conf/iclr/iclr2021.html#DosovitskiyB0WZ21.
[8]	ZHANG Y D, LIU H Y, HU Q. Transfuse: fusing transformers and CNNs for medical image segmentation[C]// The 24th International Conference on Medical Image Computing and Computer Assisted Intervention. Cham: Springer, 2021: 14-24.
[9]	CAO H, WANG Y Y, CHEN J, et al. Swin-Unet: Unet-like pure transformer for medical image segmentation[C]// European Conference on Computer Vision. Cham: Springer, 2023: 205-218.
[10]	LIU Z, LIN Y T, CAO Y, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]// 2021 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2021: 9992-10002.
[11]	GU A, DAO T. Mamba: Linear-time sequence modeling with selective state spaces[EB/OL]. (2023-12-01) [2024-08-04]. https://arxiv.org/abs/2312.00752.
[12]	MA J, LI F, WANG B. U-mamba: enhancing long-range dependency for biomedical image segmentation[EB/OL]. (2024-01-09) [2024-08-04]. https://arxiv.org/abs/2401.04722.
[13]	LIU Y, TIAN Y J, ZHAO Y Z, et al. VMamba: visual state space model[EB/OL]. (2024-01-18) [2024-08-04]. https://proceedings.neurips.cc//paper_files/paper/2024/hash/baa2da9ae4bfed26520bb61d259a3653-Abstract-Conference.html.
[14]	RUAN J C, XIANG S C. VM-UNet: vision mamba UNet for medical image segmentation[EB/OL]. (2024-02-04) [2024-08-31]. https://arxiv.org/abs/2402.02491v1.
[15]	ZHANG M Y, YU Y, JIN S, et al. VM-UNET-V2: rethinking vision mamba UNet for medical image segmentation[C]// The 20th International Symposium on Bioinformatics Research and Applications. Cham: Springer, 2024: 335-346.
[16]	WANG Z Y, ZHENG J Q, ZHANG Y C, et al. Mamba-UNet: UNet like pure visual mamba for medical image segmentation[EB/OL]. (2024-02-07) [2024-08-31]. https://arxiv.org/abs/2402.05079.
[17]	LIAO W B, ZHU Y H, WANG X Y, et al. LightM-UNet: mamba assists in lightweight UNet for medical image segmentation[EB/OL]. (2024-03-08) [2024-08-31]. https://arxiv.org/abs/2403.05246.
[18]	KALMAN R E. A new approach to linear filtering and prediction problems[J]. Journal of Basic Engineering, 1960, 82(1): 35-45. DOI URL
[19]	GU A, GOEL K, RÉ C. Efficiently modeling long sequences with structured state spaces[EB/OL]. (2021-10-31) [2024-08-31]. https://dblp.uni-trier.de/db/conf/iclr/iclr2022.html#GuGR22.
[20]	GU A, DAO T, ERMON S, et al. HiPPO: recurrent memory with optimal polynomial projections[C]// The 34th International Conference on Neural Information Processing Systems. New York: ACM, 2020: 125.
[21]	RUAN J C, XIANG S C, XIE M Y, et al. MALUNet: a multi-attention and light-weight UNet for skin lesion segmentation[C]// 2022 IEEE International Conference on Bioinformatics and Biomedicine. New York: IEEE Press, 2022: 1150-1156.
[22]	CODELLA N C F, GUTMAN D, CELEBI M E, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (ISBI), hosted by the international skin imaging collaboration (ISIC)[C]// The 15th IEEE International symposium on biomedical imaging. New York: IEEE Press, 2018: 168-172.
[23]	CODELLA N, ROTEMBERG V, TSCHANDL P, et al. Skin lesion analysis toward melanoma detection 2018:A challenge hosted by the international skin imaging collaboration (ISIC)[EB/OL]. (2019-02-09) [2024-06-04]. https://arxiv.org/abs/1902.03368.