his paper proposes a novel medical image segmentation framework — Anatomy-Guided Implicit Representation Attention Network (AIRA-Net) — designed to address the challenges posed by the complex and variable anatomical structures of the nasal cavity and its subregions. AIRA-Net leverages an implicit neural representation to extract global geometric features and employs a dedicated cross-attention module to effectively fuse multi-scale local and global features. Furthermore, a boundary-weighted loss function based on anatomical priors is integrated to enhance segmentation precision in regions with sparse features, particularly at the cavity boundaries. Extensive experiments on a dataset comprising 128 3D head CT volumes demonstrate that AIRA-Net achieves a DSC of 91.66% in nasal cavity segmentation, surpassing the second-best method nnU-Net by 4.5 percentage points. Additionally, AIRA-Net attains a HD95 of 10.75 mm, which is 2.82 mm lower than that of the second-best method U_a-Net.

medical image segmentation, nasal cavity segmentation, implicit neural representation, cross-attention mechanism, anatomy-guided prior

针对鼻腔及其细微解剖结构复杂多变、分割难度高的问题, 提出一种新颖的医学图像分割方法——解剖引导的隐式神经表示注意力网络AIRA-Net。该方法通过隐式神经表示机制提取全局几何特征, 并通过交叉注意力模块实现跨尺度特征的高效融合, 从而充分整合全局先验与局部细节信息。此外, 针对鼻腔结构中灰度相近、特征稀疏的端面区域, 提出基于解剖先验的边缘加权损失函数, 可以有效地强化分割网络对边界区域的精准定位。针对包含128例3D头部CT数据集的实验结果表明, AIRA-Net在鼻腔分割中实现 DSC为91.66%, 比次优方法nnU-Net提升4.5个百分点; HD95为10.75 mm, 比次优方法U_a-Net降低2.82 mm。