SABMamba: Spatial Anatomy-Based Module

• SABMamba is a neural module that integrates explicit spatial anatomical priors with multi-directional state-space scanning to preserve 3D image topology.
• It employs directional scans, gated convolutions, and residual integration to maintain spatial context and enhance segmentation accuracy.
• Empirical results show improved Dice scores, reduced computational complexity, and enhanced adaptability for high-dimensional medical imaging tasks.

Spatial Anatomy-Based Mamba Module (SABMamba) is a class of neural architectural modules designed to integrate explicit spatial anatomical priors, multi-directional state-space modeling, and efficient global context aggregation—principally for high-dimensional medical image analysis and synthesis tasks. SABMamba generalizes the Mamba state-space sequence architecture, replacing self-attention with selective state-space modeling, and extends it to 2D and 3D tensors while preserving anatomical topology via anatomically aligned scanning, gating, and shape-prior fusion.

1. Module Definition and Core Components

SABMamba is defined by its fusion of state-space sequence modeling (typically derived from the “Mamba” architecture) and explicit spatial anatomical priors or scan strategies, engineered to maintain or enhance coherence in three-dimensional medical images. It comprises three recurring technical motifs:

• Directional or Anatomically-Aligned State-Space Scans: Instead of flattening input volumes into arbitrary 1D sequences, SABMamba preserves spatial adjacency by scanning along anatomical or principal axes, e.g., sagittal (W), coronal (H), and axial (D) planes, or by employing spiral, bidirectional, or quadri-directional traversals (Ji, 5 Jun 2025, Zeng et al., 17 Aug 2025, Yuan et al., 12 May 2025).
• Residual or Gated Integration: Outputs of state-space branches are merged—often additively—with shortcut connections or through learned gating operations, reinforcing gradient flow and feature refinement.
• Anatomical Priors or Shape Prior Modules: In certain implementations, notably for vertebral/spinal structures, SABMamba maintains explicit trainable templates for each anatomical region, integrating them at multiple scales for regularization or direct feature fusion (Zhang et al., 2024).

Across applications, SABMamba blocks usually employ pointwise (1×1×1) and depthwise-separable (3×3×3) convolutions, SiLU or LeakyReLU activations, and layer normalization over the channel axis.

2. Directional State-Space Modeling and Multi-Plane Context

Conventional state-space models (SSMs) or Transformer-based modules for vision tasks often flatten images or volumes into long sequences, sacrificing explicit adjacency crucial for medical images. SABMamba addresses this by employing directional or anatomical scanning:

• Quadri-Directional Scans: DM-SegNet deploys four parallel Mamba units, scanning a tensor $z \in \mathbb{R}^{C\times H\times W\times D}$ in (i) forward axial (H), (ii) reverse axial, (iii) forward sagittal (W), and (iv) reverse sagittal directions. These maintain topological neighborhood relations inherent in anisotropic or isotropic medical volumes (Ji, 5 Jun 2025).
• Plane-Wise Selective Scans: SRMA-Mamba processes input volumes along the three principal anatomical planes, fusing the outputs via an “Anatomy-Based Selective Scan” (ABSS). Each plane is unfolded into 1D sequences and passed through state-space Mamba (“S6”) modules in both scan directions, before re-folding and summing the features. This delivers multi-plane, global 3D context to every voxel, critical for heterogeneously shaped structures such as the liver under cirrhotic changes (Zeng et al., 17 Aug 2025).
• Spiral and Bidirectional Sequences: ABS-Mamba formulates state propagation as spiral traversals across 2D slices and volumes, exploiting both forward and reversed scan orderings. Each sequence is processed by independent, bidirectional Mamba units with time-dependent, learnable state transition matrices, and the features are fused via residual projections (Yuan et al., 12 May 2025).

The directional scheme chosen depends on anatomical structure, modality, and target resolution. All approaches retain $O(NC)$ complexity ($N$ = number of voxels), circumventing the quadratic scaling of dense attention.

3. Integration of Anatomical Priors and Shape-Aware Modules

Certain SABMamba realizations inject explicit anatomical knowledge through shape priors:

• Vertebral Shape Prior Module (VSP): In 3D spinal segmentation, each anatomical region (e.g., T1–T12, L1–L5, S1) is assigned a trainable template $S_i \in \mathbb{R}^{h\times w\times l}$, learned jointly with the rest of the network. The VSP is split into global and local submodules, operating at multiple U-Net scales. It imposes both $L_2$ and Dice-based regularization to align predictions and priors, guiding boundary delineation and disambiguation between adjacent vertebrae (Zhang et al., 2024).
• Anatomy-Based Semantic Fusion: ABS-Mamba leverages pretrained encoders (e.g., SAM2-Hiera) to provide organ-scale semantic features, fusing these with modality-adaptive convolutional features via gating and robust fusion layers before projecting them into bidirectional Mamba blocks (Yuan et al., 12 May 2025).

The rationale is that anatomy-aware priors reduce class confusion, enhance localization under low-contrast scenarios, and mitigate modality-specific artifacts, especially in structures with repetitive or ambiguous appearance.

4. Gated Convolutions and Residual Structures

Gated spatial convolutions (GSC) precede the state-space stage in several SABMamba implementations, restoring inductive bias removed by sequence flattening:

• The GSC consists of spatial convolutional tensors, instance normalization, and channel-wise gating, producing an amplified residual feature as $GSC(z) = z + C(z) \odot G(z)_2$. Here, $C(z)$ encodes local context, $G(z)_2$ learns gating masks, and the output is stabilized through residual addition (Ji, 5 Jun 2025).
• Mamba or SSM paths are typically embedded within a residual block, reinforcing stable optimization and integrating shallow gating proposals, as in the Residual Visual Mamba Layer (Zhang et al., 2024).

Gated and residual designs help retain low-level features and facilitate information flow, particularly in high-depth networks or multi-scale skip connections.

5. Multi-Scale and Hierarchical Decoder Fusion

SABMamba blocks are not limited to encoders; they are integrated into decoders to propagate multi-scale, contextually enriched features:

• In DM-SegNet, SABMamba blocks receive fused encoder features at different scales, further process them through multi-scale fusion modules (“MMSFuE”), and propagate them via transposed convolutions. Decoder skip connections are enriched by residual addition of SABMamba outputs, enabling bidirectional state synchronization across the U-Net hierarchy (Ji, 5 Jun 2025).
• SRMA-Mamba passes encoder-stage outputs (each refined by SABMamba) to decoder-side reverse-attention modules that further enhance segmentation maps via coarse-to-fine refinement (Zeng et al., 17 Aug 2025).
• In SABMamba+VSP designs, skip connections are replaced with feature-gated fused outputs from the shape prior and Mamba branches, where skip-features and anatomical priors are adaptively gated before entering the decoding path (Zhang et al., 2024).

These strategies maximize information retention and facilitate coarse-to-fine, anatomy-sensitive predictions.

6. Quantitative Performance and Empirical Insights

SABMamba modules yield measurable improvements across multiple medical imaging tasks.

Task/Dataset	SABMamba Variant	Metric/Result	Reference
3D Spine Segmentation	SABMamba+VSP	CT Dice: 94.40% (↑0.52 vs. SPM)	(Zhang et al., 2024)
		MR Dice: 86.95% (↑0.12 vs. SPM)
Liver MRI Segmentation	SABMamba in SRMA-Mamba	Removal ⇒ −12% Dice	(Zeng et al., 17 Aug 2025)
Abdominal Synapse	Quadri-scan SABMamba	DSC: 85.44% (vs. 85.09 baseline)	(Ji, 5 Jun 2025)
Brain BraTS2023		DSC: 90.22%
MRI→CT Synthesis	Spiral SABMamba (ABS)	PSNR: ↑2dB, SSIM: ↑0.02 vs. best prev.	(Yuan et al., 12 May 2025)

Elimination or ablation of SABMamba consistently degrades both Dice scores and boundary accuracy, with especially pronounced effects in thin or ambiguously bounded structures. Gated spatial convolutions and multi-directional scans act synergistically, as confirmed by up to +1.73% DSC and −43.7% HD95 improvement over single-direction SSM baselines (Ji, 5 Jun 2025).

7. Complexity, Adaptability, and Outlook

The computational complexity of SABMamba is $O(CN)$, remaining linear in the voxel count due to directional 1D state-space scans and avoidance of dense attention. Model sizes and resource consumption are notably lower than transformer baselines (e.g., 17.2M parameters, 149GMac in SRMA-Mamba) (Zeng et al., 17 Aug 2025).

Cross-modality and cross-organ adaptability is achieved via modularized SABMamba blocks and parameter-efficient LoRA+ fine-tuning, enabling rapid domain adaptation with only a small fraction of weights updated, as exemplified in ABS-Mamba (Yuan et al., 12 May 2025).

A plausible implication is that future extensions may generalize SABMamba modules for even broader imaging, multi-modality synthesis, or topologically complex anatomical structures, given the consistent empirical benefits and architectural flexibility demonstrated across segmentation and translation domains.