ADA-3D: adaptive dimensional attention with dynamic fusion for 3D medical image segmentation

Abstract

3D medical image segmentation is frequently challenged by voxel anisotropy, low-contrast boundaries, and discontinuous inter-slice structures. This paper proposes Adaptive Dimensional Attention (ADA), a lightweight plug-and-play attention module for volumetric segmentation. ADA decouples feature calibration into hierarchical channel attention, anisotropic spatial attention, and slice-aware context attention. It further introduces a content-aware dynamic fusion mechanism (DFM) to adaptively balance these dimensions for different anatomical structures. By being integrated into nnFormer after patch embedding, ADA performs early feature calibration before Transformer encoding, enabling the subsequent encoder to receive features with stronger anatomical consistency. Experiments on ACDC cardiac MRI and COVID-19 CT segmentation show that ADA consistently outperforms representative attention modules including SE, CBAM, ECA, and EMA. Ablation results further verify that DFM and anisotropic spatial modeling are the key contributors to the performance gain.