Three-dimensional medical image segmentation requires models that can simultaneously preserve fine anatomical boundaries and represent longrange spatial dependencies. CT and MRI volumes contain large numbers of voxels and may exhibit substantial anisotropy, making memory consumption and computational efficiency important architectural considerations. Convolutional neural networks provide strong local spatial priors, whereas Transformer architectures use self-attention to establish global interactions between tokens. However, global attention has quadratic dependence on sequence length. This limitation has motivated hierarchical windowed Transformers, large-kernel convolutional networks, state-space models (SSMs), and medical foundation-model adaptation [4-15]. This review critically examines sixteen representative works spanning convolutional, Transformer, state-space, foundation-model, generative, and reproducibility paradigms: PGSAM, limited-interval classifier-free guidance, Snakemake, 3D U-Net, 3D UX-Net, MedNeXt, Swin UNETR, UNETR, Mamba-UNet, nnU-Net, ConvNeXt, TransUNet, MedSAM, VM-UNet, Mamba-SAM, and the Mamba-2/structured state-space duality framework [1-16]. The comparison focuses on representation construction, receptive-field behavior, computational complexity, volumetric modeling, information compression, transferability, and failure modes. The evidence does not support a universal winner. Instead, the literature suggests that efficient volumetric segmentation benefits from combining local boundary-sensitive representations with appropriately compressed global context and, where appropriate, pretrained semantic priors. State-space models provide an attractive alternative to explicit pairwise attention, but their practical effectiveness depends on sequence construction, representation capacity, and implementation. The review therefore frames CNNs, Transformers, SSMs, and foundation models as complementary components within a broader architectural design space.
3D Medical Image Segmentation, Vision Transformers, Mamba, ConvNeXt, Foundation Models