Introduction
Head and neck abscesses are severe, localized infections that can lead to life-threatening complications like airway obstruction or sepsis if not treated promptly. Contrast-Enhanced Computed Tomography (CECT) is the gold standard for diagnosing these lesions and planning surgical drainage. However, segmenting these abscesses is notoriously difficult—even for expert clinicians—due to complex surrounding anatomy, irregular boundaries, and heterogeneous internal fluid density.
While AI has advanced medical image segmentation, most public head and neck datasets focus on oncology. This paper introduces AbscessHeNe, the first publicly available dataset dedicated to head and neck abscess segmentation, laying the groundwork for automated diagnostic and retrieval tools.
The AbscessHeNe Dataset
The AbscessHeNe dataset comprises 4,926 CECT slices from 47 patients, featuring high-quality, pixel-wise ground truth masks. To ensure anatomical accuracy, the authors developed a rigorous, multi-stage annotation pipeline involving independent clinicians and senior expert adjudication.
Methodology
The authors benchmarked a spectrum of modern segmentation architectures, including CNNs (U-Net), Transformers (TransUNet, UNETR, Swin-UNETR), and State Space Models (Swin-UMamba, VM-UNet). Beyond segmentation, the study proposes a prototype Content-Based Image Retrieval (CBIR) pipeline. In this system, segmented masks act as semantic anchors. By combining image embeddings with clinical metadata (e.g., location, multiloculated status), the system allows clinicians to query the database for visually and clinically similar historical cases to aid decision-making.
Experimental Results
Benchmarking results highlight the extreme difficulty of abscess segmentation. The best-performing model, VM-UNet, achieved a Dice Similarity Coefficient (DSC) of only 0.39, indicating significant room for algorithmic innovation. Interestingly, the classic U-Net remained highly competitive, outperforming several complex Transformer-based models.
VM-UNet demonstrated superior computational efficiency, requiring only 4.11 G FLOPs compared to U-Net's 54.74 G, while achieving the highest scores across DSC, mIoU, and HD95 metrics. Qualitative analysis shows that most models struggle with the fuzzy boundaries and low contrast typical of these infectious lesions.
Discussion and Future Work
The modest baseline scores confirm that standard off-the-shelf models struggle with the indistinct boundaries and class imbalance of abscesses. However, the efficiency and accuracy of Mamba-based models (like VM-UNet) suggest that State Space Models are a viable path forward for complex medical vision tasks.
Future work will focus on expanding the dataset and exploring boundary-aware loss functions and retrieval-augmented training to improve segmentation performance and clinical utility.
Summary updated