LCM-UNet: Scale-Selective Mamba with Reparameterizable Multi-Basis Local Modeling for Lightweight Skin Lesion Segmentation

Md Shihabul Islam Shovo; Nishi Kanta Paul; Dip Kumar Saha; Kishor Morol

cs.CVNov 3, 2026

LCM-UNet: Scale-Selective Mamba with Reparameterizable Multi-Basis Local Modeling for Lightweight Skin Lesion Segmentation

Authors: Md Shihabul Islam Shovo, Nishi Kanta Paul, Dip Kumar Saha, Kishor Morol

Organizations: NIMISHES Lab, Dhaka, Bangladesh · Stamford University Bangladesh, Dhaka, Bangladesh · Elite Research Lab, New York, NY, USA

Abstract

Accurate lesion segmentation underpins automated dermoscopic analysis: masks enable measurements of asymmetry, border and size. Deployment in primary care or on mobile devices requires a compact network. Existing ultra-light Mamba U-Nets offer global context, but global-modeling cost depends on each Mamba block's resolution, while their local branches rely on a single square kernel. We introduce LCM-UNet, a 33,474-parameter U-shaped network that fuses every skip at its native resolution, restricts bidirectional Mamba to stages at one-sixteenth of the input resolution or coarser, and uses local convolutions at finer scales. A multi-basis reparameterizable adapter augments each local branch during training with dilated, horizontal and vertical depthwise kernels, then folds them into the host kernel before inference. Each skip connection also learns a stage-specific gain. Nine networks were retrained on the official ISIC 2017 and ISIC 2018 partitions under a harmonized protocol, with PH $^2$ reserved for external validation and ablations run on a development split. LCM-UNet achieved Dice scores of 88.62%, 89.73% and 90.46% on ISIC 2017, ISIC 2018 and PH $^2$ , respectively, ranking first on both ISIC test sets and second on PH $^2$ among the compared models. In the native-resolution host, removing Mamba from the one-eighth-resolution decoder stage eliminated 63% of Mamba operations, corresponding to 5.3% of all operations, with a Dice change of $-0.03$ percentage points. The adapter improved Dice by 0.48 percentage points over a single host kernel without adding deployed parameters or operations. With latency so far measured only on a GPU, LCM-UNet is a candidate for resource-constrained deployment.

Figures & tables

Fig. 1: Overall architecture of LCM-UNet. Bidirectional Mamba is used only in the shaded stages, and the output shown is the predicted lesion mask.

Panel of Figure 2 — Fig. 2: Building blocks of LCM-UNet. (a) GLMBP block with a shared bidirectional Mamba. (b) MBRA before and after reparameterization.

Dataset	Role	Train	Val.	Test	Total
ISIC 2017 [ 36 ]	Benchmark	2,000	150	600	2,750
ISIC 2018 [ 37 , 38 ]	Benchmark	2,594	100	1,000	3,694
ISIC 2018 development	Ablation	2,075	100	519	–
PH 2 [ 39 ]	External validation	–	–	200	200

TABLE I: Datasets and partitions. The ISIC 2018 development split is held out from its training set.

Method	Year	Params (M)	GMACs	DSC (%)	IoU (%)	SE (%)	SP (%)	DSC (%)	IoU (%)	SE (%)	SP (%)
				ISIC 2017 (600 test images)				ISIC 2018 (1,000 test images)
U-Net [ 4 ]	2015	17.263	39.980	86.34 $\pm$ 0.41	75.96 $\pm$ 0.63	85.12	95.85	88.21 $\pm$ 0.35	78.91 $\pm$ 0.56	87.93	96.37
UNeXt [ 13 ]	2022	1.472	0.562	86.92 $\pm$ 0.33	76.87 $\pm$ 0.52	86.05	95.89	88.54 $\pm$ 0.29	79.44 $\pm$ 0.47	88.36	96.43
MALUNet [ 8 ]	2022	0.178	0.083	87.35 $\pm$ 0.36	77.54 $\pm$ 0.57	86.41	96.06	88.73 $\pm$ 0.31	79.74 $\pm$ 0.50	88.18	96.64
EGE-UNet [ 9 ]	2023	0.053	0.072	87.52 $\pm$ 0.31	77.81 $\pm$ 0.49	86.93	95.97	88.86 $\pm$ 0.26	79.95 $\pm$ 0.42	88.91	96.44
VM-UNet [ 7 ]	2025	27.428	5.049	87.81 $\pm$ 0.29	78.27 $\pm$ 0.46	88.47	95.52	89.37 $\pm$ 0.22	80.78 $\pm$ 0.36	89.64	96.52

TABLE II: Comparison on the ISIC 2017 and ISIC 2018 test sets (mean

\pm

std over three seeds). Bold: best; underline: second best.

Method	DSC (%)	IoU (%)	SE (%)	SP (%)
U-Net [ 4 ]	88.93 $\pm$ 0.52	80.07 $\pm$ 0.84	91.84	95.07
UNeXt [ 13 ]	89.47 $\pm$ 0.47	80.95 $\pm$ 0.77	92.26	95.32
MALUNet [ 8 ]	89.62 $\pm$ 0.49	81.19 $\pm$ 0.80	91.95	95.56
EGE-UNet [ 9 ]	89.81 $\pm$ 0.41	81.50 $\pm$ 0.68	92.40	95.52
VM-UNet [ 7 ]	90.71 $\pm$ 0.33	83.00 $\pm$ 0.55	93.12	95.91
UltraLight VM-UNet [ 10 ]	89.74 $\pm$ 0.44	81.39 $\pm$ 0.72	92.18	95.56

TABLE III: External validation on PH 2 with the three ISIC 2018 models of each method. Bold: best; underline: second best.

Fig. 3: Qualitative comparison of UltraLBM-UNet and LCM-UNet on ISIC 2018 (a)–(e) and PH 2 (f). Values are per-image DSC (%), and the bottom row magnifies the dashed boxes.

(a) Component ablation
ID	$\kappa$	Adapter	D2 Mamba	Params (M)	GMACs	MMACs	DSC (%)	IoU (%)
A $-$ 1	shared	none	✓	0.0342	0.062	4.4	88.15 $\pm$ 0.29	78.81 $\pm$ 0.46
A0	shared	none	✓	0.0347	0.208	18.9	88.35 $\pm$ 0.24	79.13 $\pm$ 0.39
A1	stage	none	✓	0.0347	0.208	18.9	88.44 $\pm$ 0.22	79.28 $\pm$ 0.35
A2	stage	RDLA	✓	0.0347	0.208	18.9	88.61 $\pm$ 0.20	79.55 $\pm$ 0.32
A3	stage	RDLA	✗	0.0335	0.197	6.9	88.58 $\pm$ 0.23	79.50 $\pm$ 0.37

TABLE IV: Ablation study on the ISIC 2018 development split (519 images). DSC and IoU are mean

\pm

std over three seeds. A

-

1 is the released UltraLBM-UNet, which runs its decoder blocks before upsampling. A

-

1 to A0 combines the schedule and implementation changes. MMACs denote the Mamba cost.

Fig. 4: Accuracy and efficiency on ISIC 2018. (a) DSC versus parameters and (b) DSC versus GMACs, both using the official test set. (c) Mamba cost per stage in MMACs, with D1 and D2 of the released UltraLBM-UNet at

8\times 8

. Removing the D2 Mamba layer of the host cuts its Mamba operations by 63% but its total GMACs by only 5.3%.

Model	Params	GMACs	ms	FPS	DSC	IoU	$\Delta_{\max}$
Host (A0)	34,720	0.208	7.21	138.7	–	–	–
LCM-UNet (unfused)	34,014	0.197	6.69	149.5	89.73	81.37	–
LCM-UNet (deployed)	33,474	0.197	6.43	155.5	89.73	81.37	$2.4\times 10^{-6}$

TABLE V: Deployment efficiency and fusion check on the ISIC 2018 test set.

\Delta_{\max}

is the largest logit difference between the unfused and deployed models.

Structural Reparameterization

See all 8 papers

MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation
Apr 22, 2026Md Maklachur Rahman, Soon Ki Jung, Tracy HammondMamba Multi-Scale Feature Fusion
Texas A&M University, College Station, TX, USA · Kyungpook National University, Daegu, South Korea
LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder
Sep 23, 2026Andrei Arhire, Mihaela-Elena Breabăn, Radu TimofteMedical Image Segmentation Lightweight CNNs
Faculty of Computer Science Alexandru Ioan Cuza University of Iasi · Computer Vision Lab, CAIDAS & IFI University of Wurzburg, Germany
EA-LiteUNet: An Edge-Adaptive and Resource-Efficient U-Net for Boundary-Sensitive Dermoscopic Image Segmentation
Aug 16, 2026Wang Jiangtao, Nur Intan Raihana Ruhaiyem, Fu Panpan +2Image Segmentation Skin Lesion Segmentation
School of Network Communication, Zhejiang Yuexiu University, Shaoxing,312000, China · School of Computer Sciences, Universiti Sains Malaysia, Penang,11800, Malaysia

Apr 22, 2026J/K move · Enter open · S save

MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation

Md Maklachur Rahman, Soon Ki Jung, Tracy Hammond

Texas A&M University, College Station, TX, USA · Kyungpook National University, Daegu, South Korea

Mamba Multi-Scale Feature Fusion Medical Image Segmentation Skin Lesion Segmentation

Open paper AskPDF arXiv

Recent segmentation models have demonstrated promising efficiency by aggressively reducing parameter counts and computational complexity. However, these models often struggle to accurately delineate fine lesion boundaries and texture patterns essential for early skin cancer diagnosis and treatment planning. In this paper, we propose MambaLiteUNet, a compact yet robust segmentation framework that integrates Mamba state space modeling into a U-Net architecture, along with three key modules: Adaptive Multi-Branch Mamba Feature Fusion (AMF), Local-Global Feature Mixing (LGFM), and Cross-Gated Attention (CGA). These modules are designed to enhance local-global feature interaction, preserve spatial details, and improve the quality of skip connections. MambaLiteUNet achieves an average IoU of 87.12% and average Dice score of 93.09% across ISIC2017, ISIC2018, HAM10000, and PH2 benchmarks, outperforming state-of-the-art models. Compared to U-Net, our model improves average IoU and Dice by 7.72 and 4.61 points, respectively, while reducing parameters by 93.6% and GFLOPs by 97.6%. Additionally, in domain generalization with six unseen lesion categories, MambaLiteUNet achieves 77.61% IoU and 87.23% Dice, performing best among all evaluated models. Our extensive experiments demonstrate that MambaLiteUNet achieves a strong balance between accuracy and efficiency, making it a competitive and practical solution for dermatological image segmentation. Our code is publicly available at: https://github.com/maklachur/MambaLiteUNet.

Abstract

Structural Reparameterization

Explore similar work

MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation

LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder

EA-LiteUNet: An Edge-Adaptive and Resource-Efficient U-Net for Boundary-Sensitive Dermoscopic Image Segmentation