Abstract
The advent of Transformer- and Mamba-based architectures has significantly advanced 3D medical image segmentation by enabling long-range contextual modeling, a capability traditionally limited in convolutional neural networks (CNNs). However, state-of-the-art Transformer models often incur substantial computational complexity and parameter overhead, which becomes particularly prohibitive for volumetric inputs and is further exacerbated by the limited availability of annotated medical imaging data. To address these limitations, this work introduces SegMaFormer , a lightweight stage-aware hybrid architecture that integrates Mamba, Gated Hybrid Conv-SSM (GHConv-SSM), and efficient self-attention within a hierarchical volumetric encoder. Specifically, SegMaFormer employs a Mamba block at the first high-resolution stage to capture dense spatial dependencies with reduced overhead, uses GHConv-SSM blocks at intermediate stages to adaptively fuse state-space contextual modeling with local convolutional priors, and reserves self-attention for the final low-resolution stage to refine compact semantic representations. This design is further enhanced with generalized rotary position embeddings to improve spatial awareness in 3D feature learning. Despite its compact structure, SegMaFormer achieves competitive performance on three public benchmarks, including BTCV, BraTS, and ACDC, while using substantially fewer parameters and lower computational cost than larger Transformer- and Mamba-based counterparts. These results demonstrate that stage-aware allocation of Mamba, gated Conv-SSM fusion, and attention provides an effective accuracy-efficiency trade-off for 3D medical image segmentation.