TL;DR
This paper introduces MasHeNe, a new benchmark dataset for head and neck mass segmentation that includes both benign cysts and tumors to combat the malignancy bias of existing datasets. To tackle this challenging task, the authors propose WEMF, a U-shaped Mamba-based model that leverages tri-window CT enhancement and multi-frequency attention to significantly improve boundary detection and segmentation accuracy.
What Problem Does This Paper Address?
Head and neck masses are space-occupying lesions that can cause severe clinical complications, such as compressing the airway or esophagus and damaging adjacent nerves and blood vessels. While deep learning has advanced medical image segmentation, existing public datasets (like RADCURE or SegRap2023) focus almost exclusively on malignant tumors.
In real clinical practice, doctors encounter a mix of malignant tumors, benign tumors, and cysts. Models trained only on cancers tend to overfit to specific contrast-enhancement cues and fail on non-enhancing or low-contrast benign structures. Furthermore, the dense anatomy of the neck and the low contrast around lesion borders make precise boundary delineation incredibly difficult for standard architectures.
Methods
The authors propose the Windowing-Enhanced Mamba with Frequency integration (WEMF) model, built on a U-shaped architecture using Visual State Space (VSS) Mamba blocks as its backbone.
WEMF introduces two main novel components to improve segmentation:
- Windowing-Enhanced Module: Instead of using a single standard CT window, this module processes the input slice using three clinically distinct Hounsfield Unit (HU) window settings: Default (WL/WW=25/375), Abdomen soft-tissue (40/350), and Spine soft-tissue (20/300). These are concatenated as multi-contrast channels to expose different intensity ranges and edge profiles.
- Multi-Frequency Enhancement (MFE) Module: Placed along the skip connections, this module splits features into four parallel branches. Three branches apply 2D Discrete Fourier Transforms (DFT) across different planes (spatial and channel-spatial) to capture directional and scale-aware frequency cues, while the fourth branch uses depthwise convolution (DWConv) to retain local spatial details.
Datasets
The authors introduce MasHeNe, a newly curated dataset consisting of 65 contrast-enhanced CT (CECT) cases totaling 3,779 slices. Unlike previous datasets, it includes 29 cystic cases and 36 tumor cases, providing a balanced mix of benign and malignant pathologies.
The dataset is split into:
- Training set: 50 cases (2,931 slices)
- Validation set: 5 cases (278 slices)
- Testing set: 10 cases (570 slices)
The pixel-level annotations were verified through a rigorous three-stage clinical workflow involving independent annotation, consensus review, and expert adjudication.
Key Insights and Ideas
- Moving Beyond Cancer Bias: Including cysts and benign tumors changes the feature landscape, forcing the model to learn structural boundaries rather than just relying on contrast-enhancement cues typical of malignancies.
- The Power of Multi-Windowing: Mimicking radiologists by looking at multiple HU windows simultaneously provides complementary information. For example, the spine window preserves bone-soft tissue edges, while the abdomen window enhances parenchyma.
- Frequency-Domain Fusion: Traditional skip connections simply copy spatial features, which can carry over noise. By filtering features in the frequency domain across multiple views (spatial and channel-spatial), the model can better separate subtle lesion edges from background noise before sending them to the decoder.
Key Results
WEMF outperformed standard CNN, Transformer, and Mamba baselines (including U-Net, TransUNet, and U-Mamba) on the MasHeNe dataset:
- Segmentation Accuracy: WEMF achieved a Dice Similarity Coefficient (DSC) of 70.45%, an IoU of 66.89%, and a 95th percentile Hausdorff Distance (HD95) of 5.12 mm.
- Ablation Impact: Adding the tri-windowing technique alone boosted the baseline Dice score by nearly 6% (from 63.21% to 69.17%), confirming the utility of multi-contrast inputs.
- Efficiency: WEMF maintains a favorable efficiency profile with 72.32M parameters and 12.19G FLOPs—making it substantially lighter and less memory-intensive than heavy transformer competitors like TransUNet.
Limitations and Future Work
- Dataset Scale: At 65 cases, the dataset is still relatively small and may not capture the full clinical variability of head and neck pathologies.
- Pathology Variety: Future work aims to expand the dataset to include other space-occupying conditions like abscesses and hematomas, as well as multi-center data.
- Weak Contrast Challenges: The model still struggles with extremely weak contrast and ill-defined margins. The authors suggest exploring boundary-aware loss functions, explicit small-object modeling, and adaptive post-processing to address these difficult cases.
Summary updated