MSA-SAM: A Decoder Enhancement Framework for 3D Tumor Segmentation Based on Segment Anything

Accurate tumor segmentation in 3D medical images is essential for clinical diagnosis, treatment planning, and prognosis. However, this task remains challenging due to low contrast, irregular morphology, and indistinct boundaries. Most Segment Anything Model (SAM)-based medical imaging studies focus on improving the encoder, while paying relatively little attention to the lightweight mask decoder, particularly its role in multi-scale feature interaction and boundary modeling. To address this issue, we propose a decoder-enhanced framework, termed Multi-scale Spatial Aggregation SAM (MSA-SAM), which incorporates two dedicated modules into the mask decoding stage. Specifically, a multi-scale spatial aggregation (MSA) module is integrated within the decoder to fuse multi-resolution features through isotropic convolution, an edge enhancement branch, and a gating mechanism. In addition, a spatial-channel collaborative attention (SCA) mechanism is introduced to refine the aggregated features through spatial feature refinement and explicit channel-wise recalibration. Experiments on multiple public datasets demonstrate competitive performance across different segmentation tasks. In particular, under the same three-point GT-derived prompting setting, MSA-SAM improves Dice and NSD by 9.31 and 11.85 percentage points, respectively, compared with 3DSAM-Adapter on MSD Colon Cancer. Ablation studies further validate the effectiveness of the proposed modules, while prompt sensitivity experiments show that segmentation performance does not monotonically improve as the number of point prompts increases.