基于MedSAM的高效半监督医学图像病灶分割方法

doi:10.12068/j.issn.1005-3026.2026.20259022

摘要/Abstract

摘要：

针对半监督病灶分割中教师网络性能较差，难以指导学生网络进行有效分割的问题，本文提出一种高效的半监督医学图像病灶分割方法.该方法选用特征提取能力更强的MedSAM(medical segment anything model)作为教师网络，构建基于Mamba的轻量级学生网络，通过知识蒸馏提升学生网络分割性能.针对异构网络特征对齐带来的语义失配问题，提出基于扰动一致的跨架构知识蒸馏策略，将教师特征映射到学生特征空间并对齐扰动响应，提升学生网络特征表达能力以优化分割性能.此外，针对病灶形态多样及前景背景对比度低导致的分割一致性差问题，提出基于分布的自监督损失进行优化.在多类医学图像病灶分割数据集上的实验表明，本文方法的分割性能优于现有方法，同时学生网络参数量仅为1.34 M，显著提升了模型效率.

关键词: 病灶分割, MedSAM, Mamba, 知识蒸馏, 自监督损失

Abstract:

In semi-supervised lesion segmentation， the performance of the teacher network is poor， making it difficult for it to guide the student network to perform effective segmentation. To address this issue， an efficient semi-supervised medical image lesion segmentation method was proposed， employing the medical segment anything model （MedSAM）， which exhibited superior feature extraction capabilities， as the teacher network. A lightweight student network based on Mamba was constructed， and its segmentation performance was enhanced through knowledge distillation. To address the semantic mismatch caused by feature alignment across heterogeneous networks， a perturbation-consistent cross-architecture knowledge distillation method was introduced. This approach mapped teacher features to the student feature space and aligned perturbation responses， thereby improving the student network’s feature representation ability and improving segmentation performance. Additionally， to tackle the challenges of diverse lesion morphologies and low foreground-background contrast， leading to poor segmentation consistency， a distribution-based self-supervised loss was proposed for optimization. Experiments on multiple types of medical image lesion segmentation datasets demonstrate that the proposed method in this paper outperforms existing methods in segmentation performance. Meanwhile， the student network has only 1.34 M parameters， which significantly improves the model efficiency.

Key words: lesion segmentation, MedSAM, Mamba, knowledge distillation, self-supervised loss

中图分类号:

TP 391.4

贾熹滨, 尹训洁, 范超, 杨正汉. 基于MedSAM的高效半监督医学图像病灶分割方法[J]. 东北大学学报（自然科学版）, 2026, 47(1): 1-10.

Xi-bin JIA, Xun-jie YIN, Chao FAN, Zheng-han YANG. Efficient Semi-supervised Medical Image Lesion Segmentation Method Based on MedSAM[J]. Journal of Northeastern University(Natural Science), 2026, 47(1): 1-10.

图/表 11

参考文献 31

[1]	Tarvainen A， Valpola H. Mean teachers are better role models： weight-averaged consistency targets improve semi-supervised deep learning results［C］// Proceedings of the International Conference on Learning Representations. Vancouver， 2018：6-11.
[2]	Yu L Q， Wang S J， Li X M， et al. Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmentation［C］//Medical Image Computing and Computer Assisted Intervention—MICCAI 2019. Cham： Springer， 2019： 605-613.
[3]	赵小明，石培炼，王丹丹，等. 分布感知均值教师网络的半监督医学影像分割［J］. 中国图象图形学报， 2025， 30（2）： 575-588.
	Zhao Xiao-ming， Shi Pei-lian， Wang Dan-dan， et al. Distribution-aware mean teacher networks for semi-supervised medical image segmentation［J］. Journal of Image and Graphics， 2025， 30（2）： 575-588.
[4]	Luo X D， Liao W J， Chen J N， et al. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency［C］//Medical Image Computing and Computer Assisted Intervention—MICCAI 2021. Cham： Springer， 2021： 318-329.
[5]	李飞翔，降爱莲. MSMVT：多尺度和多视图Transformer半监督医学图像分割框架［J］. 计算机工程与应用， 2025， 61（2）： 273-282.
	Li Fei-xiang， Jiang Ai-lian. MSMVT： semi-supervised framework with multi-scale and multi-view Transformer for medical image segmentation［J］. Computer Engineering and Applications， 2025， 61（2）： 273-282.
[6]	Li S L， Zhang C Y， He X M. Shape-aware semi-supervised 3D semantic segmentation for medical images［C］//Medical Image Computing and Computer Assisted Intervention—MICCAI 2020. Cham： Springer， 2020： 552-561.
[7]	Wu Y C， Ge Z Y， Zhang D H， et al. Mutual consistency learning for semi-supervised medical image segmentation［J］. Medical Image Analysis， 2022， 81： 102530.
[8]	Bai Y H， Chen D W， Li Q L， et al. Bidirectional copy-paste for semi-supervised medical image segmentation［C］//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Vancouver， 2023： 11514-11524.
[9]	Chen X K， Yuan Y H， Zeng G， et al. Semi-supervised semantic segmentation with cross pseudo supervision［C］//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Nashville， 2021： 2613-2622.
[10]	Ouali Y， Hudelot C， Tami M. Semi-supervised semantic segmentation with cross-consistency training［C］//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. Seattle，2020： 12671-12681.
[11]	徐新辉，曾智勇，林征宇. 基于一致性正则化的半监督医学图像分割方法［J］. 中国医学物理学杂志， 2025， 42（6）： 784-790.
	Xu Xin-hui， Zeng Zhi-yong， Lin Zheng-yu. Semi-supervised medical image segmentation method based on consistency regularization［J］. Chinese Journal of Medical Physics， 2025， 42（6）： 784-790.
[12]	Miao J Z， Chen C， Zhang K L， et al. Cross prompting consistency with segment anything model for semi-supervised medical image segmentation［C］//Medical Image Computing and Computer Assisted Intervention—MICCAI 2024. Cham： Springer， 2024： 167-177.
[13]	Zhang Y C， Yang J， Liu Y C， et al. SemiSAM： enhancing semi-supervised medical image segmentation via SAM-assisted consistency regularization［C］//2024 IEEE International Conference on Bioinformatics and Biomedicine （BIBM）. Lisbon， 2024： 3982-3986.
[14]	Miao J Z， Chen C， Liu F R， et al. CauSSL： causality-inspired semi-supervised learning for medical image segmentation［C］//2023 IEEE/CVF International Conference on Computer Vision （ICCV）. Paris， 2023： 21369-21380.
[15]	Wang Y C， Xiao B， Bi X L， et al. Boundary-aware prototype in semi-supervised medical image segmentation［J］. IEEE Transactions on Image Processing， 2024， 33： 5456-5467.
[16]	Ma J， He Y T， Li F F， et al. Segment anything in medical images［J］. Nature Communications， 2024， 15（1）： 654-659.
[17]	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 （ICCV）. Montreal， 2022： 9992-10002.
[18]	Hatamizadeh A， Nath V， Tang Y C， et al. Swin UNETR： swin transformers for semantic segmentation of brain tumors in MRI images［C］//Brainlesion： Glioma， Multiple Sclerosis， Stroke and Traumatic Brain Injuries. Cham： Springer， 2022： 272-284.
[19]	Gu A， Dao T. Mamba： linear-time sequence modeling with selective state spaces［EB/OL］. （2023-12-01）［2024-12-24］. .
[20]	Gao J H， Cai Y H， Zhao Z K， et al. RGU-Mamba： an U-Mamba network with region-based training optimized for domain generalization applied to myocardial scar and edema segmentation［C］//Comprehensive Analysis and Computing of Real-World Medical Images. Cham： Springer， 2025： 77-86.
[21]	Jiao J B， Liu Y， Liu Y F， et al. VMamba： visual state space model［C］//Advances in Neural Information Processing Systems. Vancouver， 2024： 103031-103063.
[22]	Wu R K， Liu Y H， Liang P C， et al. H-vmunet： high-order vision Mamba UNet for medical image segmentation［J］. Neurocomputing， 2025， 624： 129447.
[23]	Fan C， Yu H Y， Huang Y， et al. SliceMamba with neural architecture search for medical image segmentation［J］. IEEE Journal of Biomedical and Health Informatics， 2025， 29（10）： 7446-7458.
[24]	Hinton G， Vinyals O， Dean J. Distilling the knowledge in a neural network［EB/OL］. （2015-03-09）［2024-12-24］ . .
[25]	Romero A， Ballas N， Kahou S E， et al. FitNets： hints for thin deep nets［EB/OL］. （2014-12-19）［2024-12-14］. .
[26]	Sun S Q， Cheng Y， Gan Z， et al. Patient knowledge distillation for BERT model compression［EB/OL］. （2019-08-25）［2024-12-24］. .
[27]	Hao Z， Guo J， Jia D， et al. Learning efficient vision transformers via fine-grained manifold distillation［C］//Advances in Neural Information Processing Systems. New Orleans， 2022： 9164-9175.
[28]	Zagoruyko S， Komodakis N. Paying more attention to attention： improving the performance of convolutional neural networks via attention transfer［EB/OL］. （2016-12-12）［2024-12-20］. .
[29]	Hao Z， Guo J， Han K， et al. One-for-all： bridge the gap between heterogeneous architectures in knowledge distillation［C］//Advances in Neural Information Processing Systems. New Orleans， 2023： 79570-79582.
[30]	Liu Y F， Cao J J， Li B， et al. Cross-architecture knowledge distillation［J］. International Journal of Computer Vision， 2024， 132（8）： 2798-2824.
[31]	Zhai S W， Wang G T， Luo X D， et al. PA-seg： learning from point annotations for 3D medical image segmentation using contextual regularization and cross knowledge distillation［J］. IEEE Transactions on Medical Imaging， 2023， 42（8）： 2235-2246.

方法	标注比例/%	Dice/%↑	HD/像素↓
Baseline U-Net	100	80.28	15.24
	10	64.48	20.53
	20	66.01	19.31
UA-MT	10	70.12	17.65
	20	72.29	16.18
SemiSAM-MT	10	72.47	15.34
	20	74.88	13.97
SASSNet	10	70.04	19.58
	20	71.61	17.66
URPC	10	67.43	18.01
	20	70.07	15.26
MC-Net+	10	72.89	15.52
	20	73.27	13.86
BCP	10	81.48	14.01
	20	82.21	13.57
CauSSL	10	80.18	14.25
	20	81.93	13.87
BaPC	10	80.29	13.96
	20	81.47	13.08
CPC-SAM	10	81.73	12.16
	20	83.68	11.59
Ours	10	83.46	13.60
	20	85.65	12.34

方法	标注比例/%	Dice/%↑	HD/像素↓
Baseline U-Net	100	80.28	15.24
	10	64.48	20.53
	20	66.01	19.31
UA-MT	10	70.12	17.65
	20	72.29	16.18
SemiSAM-MT	10	72.47	15.34
	20	74.88	13.97
SASSNet	10	70.04	19.58
	20	71.61	17.66
URPC	10	67.43	18.01
	20	70.07	15.26
MC-Net+	10	72.89	15.52
	20	73.27	13.86
BCP	10	81.48	14.01
	20	82.21	13.57
CauSSL	10	80.18	14.25
	20	81.93	13.87
BaPC	10	80.29	13.96
	20	81.47	13.08
CPC-SAM	10	81.73	12.16
	20	83.68	11.59
Ours	10	83.46	13.60
	20	85.65	12.34

方法	标注比例/%	Dice/%↑	HD/像素↓
Baseline U-Net	100	78.57	10.24
	10	72.47	16.16
	20	74.54	14.02
UA-MT	10	79.15	15.45
	20	81.84	12.30
SemiSAM-MT	10	79.81	14.84
	20	82.05	12.43
SASSNet	10	82.93	15.75
	20	83.36	11.61
URPC	10	77.48	14.26
	20	79.69	11.84
MC-Net+	10	83.71	12.62
	20	84.24	11.04
BCP	10	85.65	12.20
	20	87.03	11.15
CauSSL	10	86.25	11.52
	20	87.56	10.36
BaPC	10	85.87	13.84
	20	87.40	11.28
CPC-SAM	10	86.15	11.79
	20	87.91	10.35
Ours	10	87.58	10.06
	20	90.32	9.84

方法	标注比例/%	Dice/%↑	HD/像素↓
Baseline U-Net	100	78.57	10.24
	10	72.47	16.16
	20	74.54	14.02
UA-MT	10	79.15	15.45
	20	81.84	12.30
SemiSAM-MT	10	79.81	14.84
	20	82.05	12.43
SASSNet	10	82.93	15.75
	20	83.36	11.61
URPC	10	77.48	14.26
	20	79.69	11.84
MC-Net+	10	83.71	12.62
	20	84.24	11.04
BCP	10	85.65	12.20
	20	87.03	11.15
CauSSL	10	86.25	11.52
	20	87.56	10.36
BaPC	10	85.87	13.84
	20	87.40	11.28
CPC-SAM	10	86.15	11.79
	20	87.91	10.35
Ours	10	87.58	10.06
	20	90.32	9.84

方法	标准比例/%	Dice/%↑	HD/像素↓
Baseline U-Net	100	83.79	6.84
	10	69.52	11.64
	20	73.27	7.01
UA-MT	10	81.39	11.55
	20	83.84	8.95
SemiSAM-MT	10	81.64	10.48
	20	83.46	9.77
SASSNet	10	79.24	10.76
	20	81.44	9.94
URPC	10	78.04	16.97
	20	79.58	13.58
MC-Net+	10	83.49	9.55
	20	84.75	9.07
BCP	10	81.32	10.12
	20	83.12	9.53
CauSSL	10	84.19	12.68
	20	85.36	10.36
BaPC	10	84.95	8.94
	20	85.73	8.01
CPC-SAM	10	83.68	10.24
	20	85.07	9.64
Ours	10	87.63	7.59
	20	88.45	6.36