Source-linked AI summary
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, Abdulkadir Sengur, Leszek Rutkowski, Yan Tian
TL;DR
FU-Mamba addresses spatial-discontinuity and frequency-imbalance challenges in dental image segmentation. It combines dynamic content-aware scanning with frequency-domain enhancement in a visual state space model, achieving a 1.1% mIoU increase on the dental segmentation dataset. Failures remain under extremely challenging conditions, affecting approximately 6.5% of samples.
Problem
Predefined scanning patterns can disrupt spatial continuity, while dental imaging artifacts attenuate high-frequency details and hinder boundary localization.
Method
FU-Mamba combines a Dynamic Mamba Block for adaptive sampling and scanning with a frequency-domain enhancement block for balancing spectral components.
Results
1.1% mIoU increase on the dental segmentation dataset is reported for FU-Mamba.
Takeaways & Limitations
The framework generates high-quality dental segmentation masks by modeling spatial continuity and enhancing frequency representations.
Takeaways & Limitations
Approximately 6.5% of samples showed noticeable segmentation degradation under extremely challenging conditions.
Abstract
from arXiv · showhide
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/
1. Introduction
FU-Mamba addresses spatial discontinuity and frequency imbalance in dental image segmentation by combining content-adaptive scanning with frequency-domain enhancement. The framework targets difficult boundaries and challenging illumination while producing improved segmentation masks.
- Motivation: Existing visual state space models flatten image patches using scanning strategies that can disrupt semantic spatial adjacency and lose structural information.Predefined scanning patterns may also limit fine-grained local-structure preservation around irregular boundaries.
- Motivation: Dental images commonly contain uneven lighting, reflective enamel and restorations, motion blur, and noise that attenuate high-frequency edges and textures.These artifacts redistribute energy toward lower frequencies, making tooth edges and gum shapes harder to detect.
- Results: Under poor lighting, FU-Mamba achieves a more balanced frequency representation and delineates the entire tooth region compared with the low-quality U-Mamba mask.Experiments on the DSD and OralVision datasets report high-quality segmentation masks.
- Framework: FU-Mamba integrates dynamic scanning and frequency-domain enhancement within a visual state space model for dental image segmentation.Its design is tailored to complex tooth boundaries, low-contrast regions, uneven illumination, and reflection-induced frequency imbalance.
- Framework: The Dynamic Mamba Block learns sampling positions and scanning sequences from input content to preserve spatial continuity and enhance structural representation.The block dynamically adjusts scanning paths to prioritize important input features.
- Framework: The frequency-domain enhancement block balances high- and low-frequency components through wavelet decomposition and spectrum pooling.Its wavelet-guided spectral pooling also emphasizes mid-frequency signals aligned with human visual sensitivity.
2. Related Work
Prior dental segmentation research advances accuracy, boundary delineation, efficiency, and robustness through architectural improvements, while vision Mamba and frequency-aware methods extend modeling capabilities. FU-Mamba combines content-adaptive dynamic sampling with frequency enhancement to address dental-specific structures and imaging artifacts.
- Dental image segmentation: Dental segmentation methods target precise tooth structures, boundaries, and contextual information for clinical analysis, but oral edges and insufficient lighting remain limitations.These challenges motivate continued development of dental-specific segmentation models.
- Dental image segmentation: UNet-based approaches improve dental segmentation through attention gates, dense skip connections, multi-scale aggregation, dilated attention, and knowledge distillation.Reported benefits include sharper boundaries, more precise feature detection, and reduced computational overhead without compromising accuracy.
- Vision Mamba: Visual Mamba methods convert images into sequences and use varied scanning strategies to model spatial dependencies and global context efficiently.Examples include bidirectional, four-directional, local-window, and deformable scanning designs.
- Vision Mamba: Many visual Mamba methods rely on predefined scanning patterns, which may limit fine-grained local-structure preservation in images with irregular boundaries.Images lack a unique causal ordering, making scan-strategy design important for visual Mamba models.
- FU-Mamba: FU-Mamba differs by combining content-adaptive dynamic sampling with frequency-domain enhancement for complex two-dimensional dental structures.The design addresses complex tooth boundaries, low contrast, uneven illumination, and reflection-induced frequency imbalance.
- Frequency-domain analysis: Frequency-aware research uses feature disentanglement, recalibration, wavelet transforms, and adaptive spectral fusion to model frequency-specific or multi-scale information.These methods seek to mitigate frequency bias and strengthen spectral cues in spatial representations.
3. Our Approach
FU-Mamba uses a UNet-based architecture that combines dynamic scanning for content-aware feature extraction with frequency enhancement to support dental image segmentation. Its design targets spatial discontinuity in fixed scanning patterns and imbalanced frequency components.
- Overall architecture: The approach follows a UNet architecture with multiple encoders, decoders, and skip connections, using DMBs in encoders and frequency enhancement blocks in skip connections.The network takes dental images as input and produces segmentation masks.
- Motivation: Existing manually designed scanning patterns can overlook continuous correlations between image patches and fail to adapt to image content.The cited discussion also describes difficulties preserving fine-grained spatial details and connecting distant regions.
- Dynamic scanning: Dynamic scanning uses an offset prediction network to adjust sampling locations, focus on key regions, and feed bilinearly interpolated features into the SSM.The DMB is integrated across multiple feature scales, while depthwise convolutions and convolutional feed-forward networks enhance local feature extraction.
- Dynamic scanning: The dynamic scanning process reorders sampled features by spatial proximity, enabling flexible focus on salient regions and smooth information aggregation.Predicted offsets are combined with initial coordinates, normalized, and constrained to valid feature-map ranges before interpolation.
- Frequency enhancement: The frequency enhancement design balances high- and low-frequency features through wavelet-guided spectral processing and reconstructs enhanced feature maps for segmentation.The framework identifies frequency enhancement blocks and their wavelet-guided spectral pooling and spectrum pooling filter components.
- Reported outcome: The paper reports that the approach significantly improves the precision and reliability of dental image segmentation and delineates the entire tooth region.The reported visual result describes a more effective equilibrium among frequency components.
4. Results
FU-Mamba was evaluated on two publicly accessible dental datasets, with ablations showing benefits from dynamic scanning and frequency enhancement. The full design improved segmentation accuracy and robustness, while remaining computationally efficient, though failures persisted under severe imaging conditions.
- Dataset evaluation: Evaluations used the DSD and Oralvision dental datasets, with comparative experiments against alternative methodologies.The DSD subset supports multi-class semantic segmentation of oral structures.
- Scanning ablation: 1.2% higher mIoU than CS on DSD shows that DMB improves scanning-based segmentation.The improvement is associated with dynamic adjustment to spatial variations, local-context modeling, and more precise boundary delineation.
- Frequency enhancement ablation: 1.4% lower mIoU without FEB demonstrates the block’s contribution, especially under noise and insufficient illumination.Visual comparisons report degraded segmentation quality after FEB removal in challenging regions.
- Frequency enhancement ablation: 92.3% mIoU and 88.9% mBIoU were achieved by the complete FEB, exceeding configurations using only DWT-Conv or SPF.DWT-Conv alone raised mIoU from 90.9% to 91.7%, while SPF alone reached 91.4%, indicating complementary benefits.
- Frequency-balancing analysis: λ = 0.7 produced the best DSD performance, reaching 92.3% mIoU; λ = 0.3 was more noise-sensitive and λ = 0.9 weakened fine boundaries.The results indicate a trade-off between high-frequency detail emphasis and low-frequency structural information.
- Failure analysis: Approximately 6.5% of DSD test samples showed noticeable degradation, mainly through boundary leakage and local structure omission.Failures were associated with severe noise, strong reflections, low contrast, dental calculus occlusion, and limitations in dynamic offsets and frequency enhancement.
5. Conclusion
FU-Mamba integrates dynamic scanning with frequency-domain enhancement for dental image segmentation. The method reports competitive performance across two dental datasets, while remaining limited under adverse imaging conditions.
- FU-Mamba integrates a dynamic scanning mechanism with a frequency enhancement block for dental image segmentation.The framework combines content-adaptive sampling with frequency-domain operations.
- Experimental results: The proposed methodology attains competitive segmentation performance across two dental datasets.
- Dynamic scanning: The Dynamic Mamba Block generates reordered sequences from adaptively adjusted positions to capture critical structures more accurately.
- Frequency enhancement: The Frequency Enhancement Block balances spectral components and improves depiction of intricate frequency characteristics.
- Limitations: Under noisy conditions, limitations in fine-grained structure capture may produce incomplete or inaccurate boundary delineation.The discussion attributes these failures to dynamic offset estimation and frequency-domain enhancement under adverse imaging conditions.
Declaration of Competing Interest
The authors declare no conflicts of interest regarding publication of the article.
- All authors declare no conflicts of interest with respect to publication of the article.
Availability of Data and Materials
Study data are available from the corresponding author upon reasonable request.
- Data supporting the study findings are available from the corresponding author on reasonable request.