TL;DR
This paper introduces AbscessHeNe, the first publicly available, comprehensively annotated contrast-enhanced CT (CECT) dataset dedicated to head and neck abscess segmentation. By benchmarking convolutional, Transformer, and state-space (Mamba) architectures, the authors demonstrate that while Mamba-based models lead the pack, the intricate anatomy and low contrast of abscess boundaries make this an exceptionally challenging task. Furthermore, the dataset is uniquely structured to pioneer future content-based image retrieval (CBIR) systems for clinical decision support.
What Problem Does This Paper Address?
Head and neck abscesses are acute, pus-filled infections that can rapidly escalate into life-threatening complications like airway obstruction or septic shock if not treated promptly. While contrast-enhanced computed tomography (CECT) is the gold standard for evaluation, precisely delineating abscess boundaries is incredibly difficult. The complex anatomy of the deep cervical spaces—packed with vital blood vessels and nerves—becomes distorted by inflammation and edema.
While AI has made massive strides in medical imaging, existing head and neck CT datasets focus almost exclusively on oncology (tumor segmentation and organs-at-risk for radiotherapy planning). This leaves a critical gap for non-oncologic, emergency conditions. The authors address this lack of data by releasing a curated dataset and establishing baseline performance metrics to spark research in automated abscess segmentation and case-based clinical retrieval.
Methods
The authors established a rigorous, multi-stage data annotation pipeline to ensure ground-truth reliability:
- Initial Annotation: Two independent doctors manually segmented abscess boundaries at the pixel level.
- Consensus Review: The annotators jointly resolved any discrepancies.
- Expert Adjudication: A senior expert reviewed all segmentations and reports to ensure anatomical accuracy and make final decisions on disputed cases.
Using this dataset, they benchmarked three distinct deep learning paradigms:
- CNNs: Traditional U-Net, which excels at local spatial details.
- Transformers: TransUNet, UNETR, and Swin-UNETR, which leverage self-attention to capture long-range dependencies.
- State Space Models (Mamba): Swin-UMamba and VM-UNet, which offer long-range modeling capabilities similar to Transformers but with highly efficient linear complexity.
Beyond standalone segmentation, the authors designed a conceptual framework for a Content-Based Image Retrieval (CBIR) prototype. This pipeline uses a vector database (FAISS) to match image embeddings with structured clinical metadata (e.g., location, patient age, multiloculated status), allowing clinicians to search for visually and contextually similar historical cases.
Datasets
The newly introduced AbscessHeNe dataset consists of 4,926 contrast-enhanced CT slices derived from 47 clinically confirmed patient cases.
- Data Split: 80% training (3,935 images) and 20% testing (991 images).
- Format: Images are provided in PNG format, resized to 512 × 512 pixels, and paired with pixel-wise ground-truth masks.
- Availability: The dataset is open-sourced on GitHub.
Key Insights and Ideas
- The Mamba Advantage: Mamba-based architectures, particularly VM-UNet, proved to be highly effective for this task. They successfully balance the global context-awareness of Transformers with a fraction of the computational overhead (FLOPs).
- A High Bar for Medical AI: The top-performing model only achieved a Dice Similarity Coefficient (DSC) of 0.39. This modest score underscores that head and neck abscess segmentation is far from solved. The irregular shapes, variable sizes, and low contrast at the soft-tissue interface present a severe challenge to current state-of-the-art models.
- Moving Beyond Segmentation: By pairing pixel-level masks with rich metadata, the authors shift the paradigm from pure segmentation toward "case-based reasoning," where an AI doesn't just draw boundaries but helps doctors find similar past cases to guide surgical drainage planning.
Key Results
Quantitatively, VM-UNet outperformed all other architectures across almost every metric while remaining incredibly lightweight:
- VM-UNet: Achieved the highest DSC of 0.39, a mean Intersection-over-Union (mIoU) of 0.27, a Normalized Surface Distance (NSD) of 0.67, and the lowest 95th percentile Hausdorff Distance (HD95) of 21.27 mm. Crucially, it required only 4.11 G FLOPs of computational power.
- U-Net: Was the strongest non-Mamba baseline, achieving a DSC of 0.35, though it required significantly more computation (54.74 G FLOPs).
- Transformers: Pure Transformer models like UNETR struggled significantly on this task, posting a lower DSC of 0.22.
Qualitative evaluations (Figure 3) confirmed that while most models suffered from false positives and struggled with indistinct boundaries, VM-UNet's predictions aligned most closely with the expert ground truth.
Limitations and Future Work
The primary limitation of this study is the cohort size: the dataset is built from 47 patients (split into 40 training and 7 testing cases), evaluated on a single train/test split. This limits immediate statistical robustness and generalizability.
To address this, the authors plan to expand AbscessHeNe with more cases and richer metadata. Future algorithmic directions will explore boundary-aware loss functions, multi-window Hounsfield Unit (HU) normalization to better capture tissue density, and 3D contextual modeling to improve boundary precision.
Summary updated