FU-Mamba: A Frequency-Enhanced Dynamic Scanning Framework for Oralscan Image Segmentation

Xinxin Zhao; Jinpeng Ye; Bo Wei; Liqin Wu; Mahmoud Hassaballah; Karen Egiazarian; Aura Conci; Victor Hugo C. de Albuquerque

cs.CVAug 27, 2026

FU-Mamba: A Frequency-Enhanced Dynamic Scanning Framework for Oralscan Image Segmentation

Authors: Xinxin Zhao, Jinpeng Ye, Bo Wei, Liqin Wu, Mahmoud Hassaballah, Karen Egiazarian, Aura Conci, Victor Hugo C. de Albuquerque,

Organizations: School of Computer Science and Technology, Zhejiang Gongshang University, 310018, Hangzhou, China

Abstract

Oralscan image segmentation is essential for computer-aided diagnosis and treatment planning in digital dentistry. However, existing visual state space models (SSMs) often rely on manually designed scanning orders to flatten image patches into sequences, which disrupts the semantic spatial continuity and hinders coherent feature extraction from key foreground regions. Moreover, elements such as inconsistent lighting, reflective surfaces, and noise during data acquisition disrupt the frequency distribution by diminishing high-frequency details while enhancing low-frequency components, consequently hindering the accurate localization of boundaries. In response to these challenges, we introduce FU-Mamba, an innovative framework that incorporates dynamic scanning and frequency domain enhancement within the SSM architecture. Specifically, the Dynamic Mamba Block (DMB) adaptively learns sampling offsets via a trainable offset prediction network and performs flexible bilinear interpolation, enabling content-aware scanning that preserves spatial coherence. Furthermore, a frequency domain enhancement block balances spectral components through wavelet-guided decomposition and spectrum pooling, improving robustness under adverse imaging conditions. Experimental findings indicate that FU-Mamba attains a notable enhancement in segmentation accuracy, evidenced by a 1.1% increase in the mean intersection over union (mIoU) metric when evaluated on the dental segmentation dataset. Project page: https://byte2bite.github.io/FU-Mamba/

Figures & tables

Figure 1: Comparison of different scanning strategies in visual state space models. (a) Continuous scanning strategy proposed by PlainMamba; (b) Local scanning strategy proposed by LocalMamba; (c) Our proposed dynamic scanning strategy; (d) Original dental image under poor lighting conditions; (e) Low-quality segmentation mask produced by U-Mamba; (f) Segmentation mask generated by our approach. Yellow dots indicate reference points, and yellow arrows represent the scanning order.

Figure 2: The comprehensive framework of the proposed approach. The proposed approach follows the UNet architecture and comprises multiple encoders, decoders, and skip connections. In particular, the skip connections are constructed using frequency domain enhancement blocks. Meanwhile, encoders are built upon DMBs, ensuring effective feature extraction and reconstruction. In the figure, FEB refers to the frequency domain enhancement block, WSPM denotes a wavelet-guided spectral pooling module, and SPF represents the spectrum pooling filter. The blue area represents the mask generated after image segmentation.

Figure 3: Illustration of the DMB, where only six reference points are displayed for clarity. Each starting reference point corresponds to a patch’s origin, with its offset calculated via the OPN. Utilizing the predicted two-dimensional coordinates, bilinear interpolation is applied to extract features from significant areas, which are subsequently input into the SSM. To demonstrate our scanning approach, the figure displays six sample points, though the full process includes additional points.

Scanning Strategy	DSD (mIoU)	OralVision (mIoU)
CS	89.7 $\pm$ 0.4	89.1 $\pm$ 0.5
LS	90.2 $\pm$ 0.3	89.8 $\pm$ 0.4
DMB	90.9 $\pm$ 0.3	90.4 $\pm$ 0.3

Table 1: Ablation study on the scanning design. CS refers to the continuous scanning strategy, LS denotes the local scanning strategy, and DMB denotes the proposed Dynamic Mamba Block. Results are reported as mean

\pm

standard deviation over three independent runs.

Panel of Figure 4 — Figure 4: Comparison of mIoU scores on the dental segmentation dataset and the OralVision dataset with and without the FEB, as well as with the FEB-w/o WSPM variant. Here, FEB-w/o WSPM refers to a configuration in which the WSPM component within FEB is replaced by DWT-Conv. As shown, the model incorporating the full frequency domain enhancement block achieves higher mIoU than the other two configurations.

Figure 5: Visualization comparison of the frequency domain enhancement block. The FEB module enables the segmentation results on low-light dental images to more closely align with the ground truth. The blue regions indicate the predicted segmentation results, while the yellow dashed boxes highlight differences across various settings, emphasizing the advantages of incorporating FEB under challenging lighting conditions.

SPF	DWT-Conv	mIoU (%)	mBIoU (%)
		90.9 $\pm$ 0.3	86.9 $\pm$ 0.2
✓		91.4 $\pm$ 0.3	87.5 $\pm$ 0.3
	✓	91.7 $\pm$ 0.2	87.9 $\pm$ 0.3
✓	✓	92.3 $\pm$ 0.2	88.9 $\pm$ 0.3

Table 2: Ablation study on the effectiveness of different frequency enhancement components in FEB.

Figure 6: Effect of different frequency balancing weights

\lambda

DMB	FEB	mIoU (%)	mBIoU (%)	Params (M)	Latency (ms)
		90.1 $\pm$ 0.3	85.8 $\pm$ 0.4	32.8	21.3
✓		90.9 $\pm$ 0.3	86.7 $\pm$ 0.3	34.1	25.7
✓	✓	92.3 $\pm$ 0.2	88.9 $\pm$ 0.3	36.2	31.8

Table 3: Incremental ablation and complexity analysis on the DSD.

Approach	DSD		OralVision
Approach	mIoU	mBIoU	mIoU	mBIoU
SAM [ 20 ]	88.7 ± 0.4	85.3 ± 0.5	88.1 ± 0.3	85.0 ± 0.4
HQ-SAM [ 19 ]	91.2 ± 0.3	88.1 ± 0.4	90.3 ± 0.3	87.2 ± 0.3
EfficientSAM [ 49 ]	87.2 ± 0.5	84.4 ± 0.5	86.9 ± 0.4	83.8 ± 0.3
GroundedSAM [ 34 ]	88.6 ± 0.4	85.5 ± 0.3	87.6 ± 0.3	84.7 ± 0.3
nnU-Net [ 17 ]	89.7 ± 0.3	86.1 ± 0.3	89.3 ± 0.3	85.6 ± 0.4

Table 4: Experimental results on the test set of DSD and the Oralvision dataset. Our experimental results are highlighted in purple.

Figure 7: Visualization results of U-Mamba, HQ-SAM, and our approach on DSD. Our approach generates finer segmentation masks compared to others. The blue area shows the segmentation mask, and the white dashed box highlights differences between our approach and the others.

Approach	Params (M)	FLOPs (G)	Latency (ms)
Vim	26.1	41.3	20.7
U-Mamba	32.8	48.6	24.9
HQ-SAM	91.4	168.2	69.5
FU-Mamba (Ours)	36.2	53.1	28.2

Table 5: Complexity comparison of different segmentation methods.

Figure 8: Failure cases of our approach in DSD. The blue regions represent the masks, and the white dashed boxes highlight segmentation differences.

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

Innovative Tooth Segmentation Using Hierarchical Features and Bidirectional Sequence Modeling
Feb 25, 2026Xinxin Zhao, Jian Jiang, Yan Tian +5Medical Image Segmentation
School of Computer Science and Technology, Zhejiang Gongshang University, 310018, Hangzhou, China · School of Computer and Cyber Sciences, Communication University of China, 100024, Beijing, China · Center of Big Data, China Digital Culture Group Co., Ltd, 100176, Beijing, China +6
MATHENA: Mamba-based Architectural Tooth Hierarchical Estimator and Holistic Evaluation Network for Anatomy
Apr 1, 2026Kyeonghun Kim, Jaehyung Park, Youngung Han +17Dental Imaging Medical Imaging
OUTTA, Seoul, South Korea · Seoul National University, Seoul, South Korea · Sangmyung University, Seoul, South Korea +4
TopoMamba: Topology-Aware Scanning and Fusion for Segmenting Heterogeneous Medical Visual Media
Apr 28, 2026Fuchen Zheng, Chengpei Xu, Long Ma +10Image Segmentation Healthcare
University of Macau · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Dalian University of Technology +4

Feb 25, 2026J/K move · Enter open · S save

Innovative Tooth Segmentation Using Hierarchical Features and Bidirectional Sequence Modeling

Xinxin Zhao, Jian Jiang, Yan Tian, Liqin Wu+4

School of Computer Science and Technology, Zhejiang Gongshang University, 310018, Hangzhou, China · School of Computer and Cyber Sciences, Communication University of China, 100024, Beijing, China · Center of Big Data, China Digital Culture Group Co., Ltd, 100176, Beijing, China+6

Medical Image Segmentation

Open paper AskPDF arXiv

Tooth image segmentation is a cornerstone of dental digitization. However, traditional image encoders relying on fixed-resolution feature maps often lead to discontinuous segmentation and poor discrimination between target regions and background, due to insufficient modeling of environmental and global context. Moreover, transformer-based self-attention introduces substantial computational overhead because of its quadratic complexity (O(n^2)), making it inefficient for high-resolution dental images. To address these challenges, we introduce a three-stage encoder with hierarchical feature representation to capture scale-adaptive information in dental images. By jointly leveraging low-level details and high-level semantics through cross-scale feature fusion, the model effectively preserves fine structural information while maintaining strong contextual awareness. Furthermore, a bidirectional sequence modeling strategy is incorporated to enhance global spatial context understanding without incurring high computational cost. We validate our method on two dental datasets, with experimental results demonstrating its superiority over existing approaches. On the OralVision dataset, our model achieves a 1.1% improvement in mean intersection over union (mIoU).

Abstract

Medical Image Segmentation

Explore similar work

Innovative Tooth Segmentation Using Hierarchical Features and Bidirectional Sequence Modeling

MATHENA: Mamba-based Architectural Tooth Hierarchical Estimator and Holistic Evaluation Network for Anatomy

TopoMamba: Topology-Aware Scanning and Fusion for Segmenting Heterogeneous Medical Visual Media