GAD-MambaUNet: Direction-Group Mamba with Gradient-Adaptive DINOv3 Distillation for Lightweight Medical Image Segmentation

Fang Wang; Huitao Li; Wenhan Chao; Zheng Zhuo; Xinxin Yang

cs.CVSep 22, 2026

GAD-MambaUNet: Direction-Group Mamba with Gradient-Adaptive DINOv3 Distillation for Lightweight Medical Image Segmentation

Authors: Fang Wang, Huitao Li, Wenhan Chao, Zheng Zhuo, Xinxin Yang

Organizations: College of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing, 102617, People’s Republic of China · School of Computer Science and Engineering, Beihang University, Beijing, 100191, 100083, People’s Republic of China

Abstract

In this paper, we proposed GAD-MambaUNet, a lightweight medical image segmentation network that combines efficient local modeling, direction--group state-space interaction, and training-time foundation-model supervision. To improve contextual modeling in compact segmentation networks, we introduced Direction-Group Graph Selective Scan (DG-GSS), which treated scan-direction and channel-group responses as graph nodes and enabled structured information exchange before multi-directional fusion. We further incorporated DINOv3-GAD supervision, where a frozen DINOv3 teacher provided semantic guidance during training, and Gradient-Adaptive Distillation dynamically regulated the distillation strength. GAD-MambaUNet achieves a favorable accuracy--efficiency balance compared with representative lightweight and general segmentation methods. Ablation studies further verify the effectiveness of DG-GSS and training-time DINOv3-GAD supervision. In future work, we will explore more flexible teacher--student alignment strategies and extend the proposed framework to more diverse medical segmentation scenarios, such as multi-class and multi-modal segmentation tasks.

Figures & tables

Figure 1: Accuracy–efficiency comparison across PH 2 , ISIC2018, CVC-ClinicDB, and CVC-ColonDB. The vertical axis reports the macro-average Dice score over the four datasets, while the horizontal axes show student-side parameters and FLOPs. Compared with existing lightweight segmentation networks, GAD-MambaUNet achieves a more favorable Pareto balance between segmentation accuracy and deployment cost. The DINOv3 teacher is used only during training and is excluded from deployment cost.

Figure 2: Overall architecture of GAD-MambaUNet. High-resolution stages retain multi-kernel local modeling, whereas deep stages use grouped state-space blocks with DG-GSS. During training, a frozen DINOv3 teacher supervises the first decoder block and GAD regulates the distillation coefficient according to its gradient share. The teacher and projection head are removed at inference.

Method	Params(M)	FLOPs(G)	Throughput(/s)	PH 2	ISIC2018	ClinicDB	ColonDB	Avg.
U-Net	34.53	65.53	92.74	93.68	86.67	91.43	83.95	88.93
PraNet	32.55	6.93	49.94	94.16	88.46	91.71	89.16	90.87
UACANet	69.16	31.51	33.72	94.37	88.72	93.29	89.76	91.53
TransUNet	105.32	38.52	50.91	94.63	89.04	93.18	89.97	91.70
UNeXt	1.47	0.57	134.33	93.60	87.78	90.20	83.84	88.86
CMUNeXt	0.418	1.09	136.68	94.02	87.51	92.82	83.85	89.55

Table 1: Quantitative comparison with representative lightweight medical image segmentation methods on PH 2 , ISIC2018, CVC-ClinicDB, and CVC-ColonDB. Dice scores are reported in %. Baseline results are reproduced or cited from their reported settings when available. Params and FLOPs measure student-side inference complexity. For GAD-MambaUNet with DINOv3 supervision, the teacher model is used only during training and is excluded from deployment cost.

Variant	Params(M)	FLOPs(G)	PH 2	ISIC2018	ClinicDB	ColonDB
MKIR baseline (MK-UNet)	0.316	0.314	94.39	88.74	93.48	90.01
w/ GroupedSS2D	0.416	0.281	94.47	88.79	93.76	90.64
w/ DG-GSS	0.493	0.322	94.90	88.91	94.04	91.24

Table 2: Ablation study of DG-GSS. Dice scores are reported in %.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 788 papers

DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation
Jul 21, 2026Mohammad Arafat Hussain, Ellen Grant, Yangming OuMamba U-Net
Boston Children’s Hospital, Harvard Medical School
USEMA: a Scalable Efficient Mamba Like Attention for Medical Image Segmentation
May 11, 2026Elisha Dayag, Nhat Thanh Tran, Jack XinImage Segmentation Hybrid CNN-Transformer Architectures
University of California Irvine, Irvine CA 92697
BiSegMamba: Efficient Bidirectional Tri-Oriented Mamba for 3D Medical Image Segmentation
May 29, 2026Bakht Zada, Chao Tong, Qile Su +13D Medical Image Segmentation State Space Models
School of Computer Science and Engineering, Beihang University. State Key Laboratory of Virtual Reality Technology and Systems, Beihang University

Jul 21, 2026J/K move · Enter open · S save

DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation

Mohammad Arafat Hussain, Ellen Grant, Yangming Ou

Boston Children’s Hospital, Harvard Medical School

Mamba U-Net BraTS 3D Medical Image Segmentation State Space Models

Open paper AskPDF arXiv

We propose parameter-efficient SSM-based U-Net architectures for 3D medical image segmentation. Convolutional U-Nets afford O(n) local mixing per layer but lack explicit global context; transformers provide global reasoning at O(n^2) cost in sequence length

n

. State-space models (SSMs), such as Mamba, offer

O(n)

global propagation per block. Yet, existing medical SSM segmenters rely on fixed scan patterns and large parameter budgets. Dynamic Adaptive Scan (DAS), which learns data-dependent reordering before selective scan, has not been applied to medical imaging or extended to 3D volumes. We propose DAMamba-UNet3D, a hybrid encoder-decoder that integrates tri-plane 3D-DAS blocks at encoder stages E2-E4 while retaining convolutions elsewhere (~5.3M parameters). On BraTS 2020 five-fold cross-validation, DAMamba-UNet3D achieves mean Dice 0.815+/-0.013 (full-volume per-case evaluation) at ~13x lower parameter cost than SegMamba (0.824+-0.014, ~70M). At comparable scale, DAMamba-L (~70M), a wide DAS-native variant with encoder-only DAMamba and a convolutional bottleneck, reaches 0.829+-0.012, surpassing retrained SegMamba by 0.5pt. Component ablations show that encoder-only DAS placement is critical as bottleneck and decoder SSM blocks lower Dice. Together, the results suggest that learned tri-plane DAS in a hybrid U-Net is competitive with, and under our large-scale design may improve upon, SegMamba's fixed Tri-orientated Mamba (ToM) scanning on BraTS 2020. Code: https://github.com/marafathussain/DAMamba-UNet3D.

Abstract

Medical Image Segmentation

Explore similar work

DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation

USEMA: a Scalable Efficient Mamba Like Attention for Medical Image Segmentation

BiSegMamba: Efficient Bidirectional Tri-Oriented Mamba for 3D Medical Image Segmentation