PMDeUNet: a Lightweight Network Model for the Efficient Segmentation of Abdominal Medical Image

doi:10.11871/jfdc.issn.2096-742X.2026.04.012

Abstract

Abstract:

[Objective] This study aims to design an efficient and lightweight abdominal medical image segmentation network model that can be deployed on clinical medical equipment. [Methods] The advantages of CNN and Mamba architectures are flexibly combined to construct a U-shaped lightweight network model PMDeUNet. A parallel V-Mamba structure is designed as the backbone of the model to reduce the model’s complexity and improve inference speed. Meanwhile, a parallel multi-scale deformable convolution module is integrated into the model to specifically enhance the segmentation accuracy of organs with complex morphologies in abdominal images. [Results] Experimental results on public datasets show that the proposed model maintains ideal average segmentation accuracy while having fewer parameters and higher inference efficiency, and the segmentation performance for complex organs is further improved. [Conclusions] The experimental results prove the effectiveness of the proposed model structure in improving the segmentation accuracy of abdominal organs and reducing model complexity.

Key words: abdominal medical image segmentation, deep learning, mamba architecture, lightweight, deformable convolution

LIN Yusong, LIANG Zhenyu, ZHAO Zhe. PMDeUNet: a Lightweight Network Model for the Efficient Segmentation of Abdominal Medical Image[J]. Frontiers of Data and Computing, 2026, 8(4): 162-178, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.04.012.

Figures/Tables 17

Fig.1

Fig.2

Fig.3

Fig.4

Fig.5

Fig.6

Table 9

Table 1

Table 2

Table 3

Fig.7

Fig.8

Table 4

Table 5

Table 6

Table 7

Table 8

References 22

[1]	ZHENG R, SHI C, WANG C, et al. Imaging-Based Staging of Hepatic Fibrosis in Patients with Hepatitis B: A Dynamic Radiomics Model Based on Gd-EOB-DTPA-Enhanced MRI[J]. Biomolecules, 2021, 11(2): 307. doi: 10.3390/biom11020307
[2]	REHMAN A, KHAN F, et al. A deep-learning based review on abdominal images[J]. Multimedia Tools and Applications, 2021, 80: 30321-30352. doi: 10.1007/s11042-020-09592-0
[3]	KAUR H, KAUR N, NEERU N. Evolution of multiorgan segmentation techniques from traditional to deep learning in abdominal CT images-a systematic review[J]. Displays, 2022, 73: 102223. doi: 10.1016/j.displa.2022.102223
[4]	CAO H, WANG Y, Chen J, et al. Swin-UNet: Unet-like pure transformer for medical image segmentation[C]. Proceedings of the European Conference on Computer Vision Workshops, Cham, Springer, 2022, 205-208.
[5]	CHEN J, LU Y, YU Q, et al. Trans-UNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers[J]. Medical Image Analysis, 2024, 97: 103208.
[6]	HUANG X, DENG Z, LI D, YUAN X. MISSFormer: An effective transformer for 2D medical image segmentation[J]. IEEE Transactions on Medical Imaging, 2023, 42(5): 1484-1494. doi: 10.1109/TMI.2022.3230943 pmid: 37015444
[7]	JUN M, LI F, WANG B. U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation[EB/OL]. 2024, https://arxiv.org/abs/2401.047-22.
[8]	LIU J, YANG H, ZHOU H Y, et al. Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining[C]. International Conference on Medical Image Computing and Computer Assisted Intervention, Springer, Marrakesh, Morocco, 2024: 1627-1636.
[9]	WANG Z, ZHENG J, et al. Mamba-UNet: UNet-Like Pure Visual Mamba for Medical Image Segmentation[EB/OL]. 2024, https://arxiv.org/abs/2402.05079.
[10]	LIAO W, ZHU Y, WANG X, et al. LightM-UNet: Mamba Assists in Lightweight U-Net for Medical Image Segmentation[EB/OL]. 2024, https://arxiv.org/abs/2403.05246.
[11]	WU R, LIU Y, NING G, et al. Ultra-Light VM-U-Net: Parallel Vision Mamba significantly reduces parameters for skin lesion segmentation[J]. Patterns, 2025, 6(11): 101298. doi: 10.1016/j.patter.2025.101298
[12]	MDMOSTAFIjUR R, MUSTAFA M, RADU M. EMCAD: efficient multi-scale convolutional attention decoding for medical image segmentation[C]. IEEE Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, 2024, 11769-11779.
[13]	ZHU X, HU H, LIN S, et al. Deformable Conv-Nets v2: More deformable, better results[C]. IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 2019, 9300-9308.
[14]	ZHANG X, ZHOU X, LIN M, et al. Shuffle-Net: An extremely efficient convolutional neural network for mobile devices[C]. IEEE Conference on Computer Vision and Pattern Recognition, Piscataway: IEEE, 2018, 6848-6856.
[15]	MARK S, ANDREW H, ZHU M, et al. MobileNetv2: Inverted residuals and linear bottlenecks[C]. IEEE Conference on Computer Vision and Pattern Recognition, Piscataway: IEEE, 2018, 4510-4520.
[16]	HU J, LI S, GANG S. Squeeze-and-Excitation Networks[C]. IEEE Conference on Computer Vision and Pattern Recognition, Piscataway: IEEE, 2018, 7132-7141.
[17]	RAHMAN M, MARCULESCU R. Multi-scale Hierarchical Vision Transformer with Cascaded Attention Decoding for Medical Image Segmentation[C]. Proceedings of the 41st International Conference on Machine Learning, 2024, 3456-3468.
[18]	JI Y, BAI H, YANG J, et al. AMOS:A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation[EB/OL]. 2022, https://arxiv.org/2206.08023.
[19]	LANDMAN B, XU Z, IGELSIAS J, et al. MICCAI multi-atlas labeling beyond the cranial vault-workshop and challenge. Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault-Workshop Challenge[M]. 2015: 5, 12.
[20]	ISENSEE F, PETERSEN J, KOHL S A, et al. nnU-Net: a self-configuring method for deep learning based biomedical image segmentation[J]. Nat Methods, 2021, 18: 203-211. doi: 10.1038/s41592-020-01008-z
[21]	PERERA S, ERZURUMLU Y, GULATI D, et al. MobileUNETR: A Lightweight End-To-End Hybrid Vision Transformer for Efficient Medical Image Segmentation[C]. Proceedings of the European Conference on Computer Vision Workshops, Springer, 2024, 281-299.
[22]	RUAN J, XIANG S, XIE M, et al. MALUNet: a multi-attention and light-weight unet for skin lesion segmentation[C]. IEEE Conference on Bioinformatics and Biomedicine (BIBM), Piscataway: IEEE, 2022, 1150-1156.

Weight Combination	Dice	HD95
ω_i=1 ($\mathit{i}\in 1,\mathit{?}2,\dots,15$)	81.65	10.92
(ω₁,??ω₂,??ω₃,ω₄) = (4, 3, 2, 1), ω_i=1 ($\mathit{i}\in 5,\mathit{?}6,\dots,15$)	81.68	10.91
(ω₁,??ω₂,??ω₃,ω₄) = (6, 3, 2, 1), ω_i=1 ($\mathit{i}\in 5,\mathit{?}6,\dots,15$)	81.68	10.91
(ω₁,??ω₂,??ω₃,ω₄) = (1, 1, 1, 1), ω_i=0.5 ($\mathit{i}\in 5,\mathit{?}6,\dots,15$)	81.64	10.92

Methods	Avg		Spleen	Kidney （R）	Kidney （L）
Methods	Dice ↑	HD95 ↓	Spleen	Kidney （R）	Kidney （L）
nnUNet^[20]	79.92	14.36	94.37	84.79	83.07
TransUNet^[5]	78.61	19.53	94.21	84.06	82.17
SwinUNet^[4]	80.21	15.91	94.69	85.50	84.61
UMamba^[7]	79.95	18.16	94.42	84.64	83.19
MambaUNet^[9]	79.45	18.85	94.39	84.13	82.97
SwinUMamba^[8]	80.98	14.75	94.65	85.56	84.78
LightMUNet^[10]	80.64	15.60	94.43	85.33	84.15
MALUNet^[22]	79.67	17.88	94.45	84.62	83.03
PMDeUNet （ours）	81.65	10.92	94.51	85.42	84.30
Methods	Gall Bladder		Esophagus	Liver	Stomach
nnUNet	69.43		78.10	95.66	76.88
TransUNet	67.50		76.45	95.33	74.59
SwinUNet	68.05		78.85	95.64	77.24
UMamba	68.12		78.80	95.57	77.08
MambaUNet	68.09		78.07	95.38	76.94
SwinUMamba	70.05		80.18	95.72	78.75
LightMUNet	69.89		78.69	95.47	77.90
MALUNet	68.03		78.66	95.50	76.97
PMDeUNet (ours)	72.02		80.25	95.70	78.63
Methods	Aorta		Inferior Vena Cava	Pancreas	Adrenal Gland （R）
nnUNet	89.95		82.92	71.73	70.54
TransUNet	88.90		81.53	70.04	69.23
SwinUNet	90.70		82.87	72.02	70.35
UMamba	89.92		82.75	72.88	70.16
MambaUNet	89.55		82.29	71.16	69.93
SwinUMamba	90.94		83.66	73.21	71.47
LightMUNet	89.63		82.99	72.85	73.02
MALUNet	89.02		82.25	71.92	71.14
PMDeUNet (ours)	90.75		85.01	75.14	74.28
Methods	Adrenal Gland（L）		Duodenum	Bladder	Prostate/Uterus
nnUNet	68.92		68.60	87.46	76.30
TransUNet	67.47		66.14	86.10	75.33
SwinUNet	68.63		68.07	87.65	78.26
UMamba	68.49		67.98	87.44	77.80
MambaUNet	68.01		67.32	87.24	76.28
SwinUMamba	69.54		68.87	88.76	78.65
LightMUNet	70.15		69.33	88.03	77.72
MALUNet	68.06		67.81	87.30	76.25
PMDeUNet (ours)	71.09		70.97	88.16	78.60

Methods	Avg		Aorta	Gallbladder	Kidney（L）
Methods	Dice ↑	HD95 ↓	Aorta	Gallbladder	Kidney（L）
TransUNet^[5]	77.48	31.69	87.23	63.13	81.87
SwinUNet^[4]	79.13	21.55	85.47	66.53	83.28
MobileUNETR^[21]	79.53	20.85	86.93	66.98	83.80
MISSFormer^[6]	81.96	18.20	86.99	68.65	85.21
MambaUNet^[9]	78.95	24.90	85.36	66.40	83.26
SwinUMamba^[8]	79.28	20.90	85.95	66.02	83.62
PMDeUNet (ours)	81.33	18.16	86.03	70.80	83.66
Methods	Kidney（R）	Liver	Pancreas	Spleen	Stomach
TransUNet	77.02	94.08	55.86	85.08	75.62
SwinUNet	79.61	94.29	56.58	90.66	70.60
MobileUNETR	79.05	94.02	60.89	89.47	75.12
MISSFormer	82.00	94.41	65.67	91.92	80.81
MambaUNet	79.03	94.45	56.74	90.76	75.57
SwinUMamba	80.70	94.28	56.43	90.36	76.85
PMDeUNet (ours)	80.99	94.35	66.43	90.22	78.12

Methods	Params （M）↓	GFLOPs ↓	FPS ↑
nnUNet^[20]	34.530	36.224	30.6
TransUNet^[5]	93.231	32.304	46.9
SwinUNet^[4]	27.147	7.818	63.2
MISSFormer^[6]	35.453	9.502	43.2
UMamba^[7]	67.260	39.920	36.8
MambaUNet^[9]	11.920	4.421	57.8
SwinUMamba^[8]	22.620	15.126	50.1
MobileUNETR^[21]	3.105	1.620	58.7
LightMUNet^[10]	1.095	1.546	62.9
MALUNet^[22]	0.806	0.912	64.3
PMDeUNet (ours)	0.753	0.865	64.6

Parallelism	Dice	HD95	Params	GFLOPs
No	81.60	10.92	1.395M	0.868
Yes	81.65	10.92	0.753M	0.865