Topology-Aware Wavelet Mamba for Airway Structure Segmentation in Postoperative Recurrent Nasopharyngeal Carcinoma CT Scans

TL;DR

TopoWMamba tackles the challenging problem of postoperative airway segmentation in recurrent nasopharyngeal carcinoma by fusing wavelet-based multi-scale features with state-space sequence modeling and topology-aware guidance. The encoder incorporates a five-stage layout with Wavelet-based Mamba Blocks and the SCVSS module to capture both local detail and global structure, while the decoder fuses multi-scale features with deep supervision. Key contributions include the integration of Wavelet transforms with Mamba blocks, the introduction of the SnakeVSS component, and the NPCSegCT dataset, achieving state-of-the-art Dice scores on NPCSegCT (88.02% average) and SegRap 2023 trachea segmentation (95.26% Dice). This work advances automated, anatomically faithful airway risk assessment in postoperative NPC patients and provides a solid foundation for future airway risk prediction models.

Abstract

Nasopharyngeal carcinoma (NPC) patients often undergo radiotherapy and chemotherapy, which can lead to postoperative complications such as limited mouth opening and joint stiffness, particularly in recurrent cases that require re-surgery. These complications can affect airway function, making accurate postoperative airway risk assessment essential for managing patient care. Accurate segmentation of airway-related structures in postoperative CT scans is crucial for assessing these risks. This study introduces TopoWMamba (Topology-aware Wavelet Mamba), a novel segmentation model specifically designed to address the challenges of postoperative airway risk evaluation in recurrent NPC patients. TopoWMamba combines wavelet-based multi-scale feature extraction, state-space sequence modeling, and topology-aware modules to segment airway-related structures in CT scans robustly. By leveraging the Wavelet-based Mamba Block (WMB) for hierarchical frequency decomposition and the Snake Conv VSS (SCVSS) module to preserve anatomical continuity, TopoWMamba effectively captures both fine-grained boundaries and global structural context, crucial for accurate segmentation in complex postoperative scenarios. Through extensive testing on the NPCSegCT dataset, TopoWMamba achieves an average Dice score of 88.02%, outperforming existing models such as UNet, Attention UNet, and SwinUNet. Additionally, TopoWMamba is tested on the SegRap 2023 Challenge dataset, where it shows a significant improvement in trachea segmentation with a Dice score of 95.26%. The proposed model provides a strong foundation for automated segmentation, enabling more accurate postoperative airway risk evaluation.

Figures (11)

• Figure 1: Partial display of our NPCSegCT dataset, showcasing annotated CT scans with critical airway-related structures.
• Figure 2: (a)The architectural design of TopoWMamba. TopoWMamba is an encoder-decoder segmentation framework that employs Mamba-based modules for effective feature extraction while maintaining low-level details through residual connections. (b)The overall structure of the SCVSS. The SCVSS features three parallel branches—conventional convolution, VSS, and SnakeVSS. (c)The illustration of Wavelet-based Mamba Block (WMB). WMB utilizes a 2D discrete wavelet transform to separate feature maps into low and high-frequency components, processing them with specialized modules to enhance long-range dependencies and global context.
• Figure 3: Details of SnakeVSS and VSS structure. In this diagram, the symbol $\oplus$ represents element-wise addition. The SnakeVSS branch reorders feature patches in serpentine patterns, capturing complex curvilinear structures, while the VSS branch focuses on conventional scanning directions to extract spatial features effectively.
• Figure 4: Details of spatial and channel attention structure. The symbol $\otimes$ denotes element-wise multiplication, and $\oplus$ represents element-wise addition. This structure enhances feature representation by focusing on important spatial regions and channel-wise dependencies, allowing the model to better capture relevant information.
• Figure 5: Schematic diagram of wavelet decomposition.