Dual-Scan STA-Mamba Architectures

• Dual-Scan STA-Mamba is a framework that combines structured state-space models with dual-path scanning to capture long-range dependencies and cross-modal interactions.
• It employs complementary scan directions—such as temporal, spatial, or spectral—with fusion techniques like elementwise addition and nonlinear gating to enhance both global and local context.
• Empirical results in fields like video processing and medical imaging demonstrate its efficiency and improved performance over traditional uni-scan architectures.

Dual-Scan STA-Mamba refers broadly to a family of architectures that combine structured state-space models (SSMs), generally the Mamba variant, with dual-path or bidirectional scanning strategies across spatial, temporal, spectral, or semantic axes. These frameworks enhance modeling power for long-range dependencies, cross-modal interaction, and global/local structural regularity in domains where standard unidirectional SSMs are limited (e.g., insufficient context, lack of cross-region semantic coherence, failure to capture bidirectional phenomena). Dual-Scan STA-Mamba systems are now deployed in video understanding, time-series modeling, cross-spectral fusion, medical imaging, audio analysis, and more, with architectures tailored for application-specific axes and fusion schemes.

1. Core Architecture and Dual-Scan Principle

Dual-Scan STA-Mamba architectures use two or more structurally distinct state-space scans (“dual scans”) of the same input, often representing:

• Opposing temporal directions (forward/reverse time)
• Orthogonal spatial traversals (e.g., row/column, kinematic tree vs. full joint list)
• Spectrum, channel, or modality axes

Each scan processes the input under the selective SSM dynamics of the Mamba framework, parameterized per-sequence-position and modulated by gates, attention, or guided weights. The outputs are then fused, commonly via elementwise sum, concatenation plus projection, gating mechanisms, or attention, before feeding into detection, segmentation, or regression heads.

The canonical SSM/Mamba recurrence for input $x_t \in \mathbb{R}^d$ is

$$h_t = \bar{A}_t h_{t-1} + \bar{B}_t x_t, \;\; y_t = C_t h_t + D x_t,$$

where $\bar{A}_t, \bar{B}_t, C_t$ may be input-dependent and are typically learned using small MLPs or gating networks on $x_t$ (Zhang et al., 19 Jun 2025, Chen et al., 29 Jan 2026, Li et al., 18 Aug 2025). Dual scans can be unidirectional (left-to-right plus right-to-left), tree-structured (kinematic chain), or guided by semantic priors (boundary maps, channel groups, or mask affinity).

2. Mathematical Formulations and Algorithms

Dual-Scan STA-Mamba blocks implement two distinct SSM traversals over the same or differently ordered input sequence $X \in \mathbb{R}^{L\times C}$:

• Global scan: canonical order (e.g., temporal, spatial index, spectral index).
• Local/semantic scan: application-informed order (e.g., joint kinematic chain, channel grouping, mask-guided region).

For time or sequence modeling, bidirectional recurrences are explicit: $$\begin{aligned} \text{Forward:} & \quad h_t^{(f)} = A_f h_{t-1}^{(f)} + B_f x_t \ \text{Backward:} & \quad h_t^{(b)} = A_b h_{t+1}^{(b)} + B_b x_t \end{aligned}$$ Fused features are then constructed as $y_t = f(h_t^{(f)}, h_t^{(b)})$, with $f$ usually being concatenation, addition, or gating (Lavaud et al., 2024, Lu et al., 22 Nov 2025).

Spatial dual scans in, for example, mesh recovery reorder the features by the kinematic tree, run per-frame SSMs, and then fuse: $$O_{\mathrm{global}} = \mathrm{SSM}(X), \quad O_{\mathrm{local}} = \mathrm{SSM}(X[\text{kinematic order}])$$ Fusing via projected nonlinearities (e.g., $$O_{\mathrm{fused}} = \mathrm{SiLU}(O_\mathrm{global}) \odot O_\mathrm{local} + O_\mathrm{global}$$) is standard (Chen et al., 29 Jan 2026).

In cross-dimensional systems (e.g., CS-MUNet), one axis is spatial (boundary-aware), the other channel:

• BASM injects boundary posterior $$h_t = \bar{A}_t h_{t-1} + \bar{B}_t x_t, \;\; y_t = C_t h_t + D x_t,$$0 into SSM forget/write parameters for spatial scan.
• CMSA treats the channel axis as a sequence, partitioned into $$h_t = \bar{A}_t h_{t-1} + \bar{B}_t x_t, \;\; y_t = C_t h_t + D x_t,$$1 groups, each with a bounded SSM recurrence (Zheng et al., 20 Mar 2026).

In audio (ESTM), time and frequency axes are decoupled—SSM over time (TMamba), SSM over frequency (SMamba)—then fused after linear alignment (Ma et al., 2 Sep 2025).

3. Application Specializations

Dual-Scan STA-Mamba frameworks are instantiated for different domains by selecting semantic axes for each scan, scan-specific SSM parameterizations, gating/fusion mechanisms, and regularization or guidance:

Domain	Dual Scans Selected	Fusion/Guidance	Reference
Video 3D Pose / Mesh	Joint-order, Kinematic-tree	SiLU-Gated Conv1D Fusion	(Chen et al., 29 Jan 2026)
Channel Estimation (OFDM)	Forward/Backward Frequency	Elementwise Sum + Gate	(Luan et al., 23 Jan 2026)
Multi-Organ Segmentation	Spatial (boundary), Channel (group-wise)	Boundary-posterior, Bounded SSM	(Zheng et al., 20 Mar 2026)
Shadow Removal	Horizontal spatial, Mask-aware adaptive	Channel Cat + Conv1×1 Residual	(Li et al., 18 Aug 2025)
Anomalous Sound	Temporal (frame), Spectral (band)	Linear Summation	(Ma et al., 2 Sep 2025)
Video Action Detection	Forward/Backward Temporal, Masked SSM	Concatenation, Diagonal Mask	(Lu et al., 22 Nov 2025)
Biophysical Time-Series	Forward/Backward Trajectory	Concatenation	(Lavaud et al., 2024)

Each field adapts the scan order and fusion to match domain priors, such as anatomical boundaries, kinematic trees, cross-modal dependencies, or spectral structure.

4. Fusion Strategies, Attention, and Selective Modulation

Fusion after dual scans is a critical design differentiator:

• Elementwise addition or concatenation: Provides combined context for downstream tasks but may result in redundancy or "self-element conflict." Diagonal masking (as in DMBSS) addresses this by nullifying the main diagonal in backward scans (Lu et al., 22 Nov 2025).
• Nonlinearity-gated fusion: SiLU gating, SE-weighted sum, or softmax mixing empirically improve alignment of local and global features (Chen et al., 29 Jan 2026, Zheng et al., 20 Mar 2026).
• Boundary/semantics-aware parameter modulation: In segmentation, pixel-level boundary posteriors modulate SSM parameters (forget/write rates), enhancing anatomical fidelity (Zheng et al., 20 Mar 2026).

Attention layers may precede or follow the state-space blocks. Mamba variants sometimes use self- or cross-attention to provide fine-grained position weighting prior to SSM processing, but dual-scan models often rely primarily on SSM-based long-range modeling due to computational and memory advantages.

5. Empirical Performance and Efficiency

Reported results consistently show that Dual-Scan STA-Mamba outperforms or matches transformer-based and uni-scan SSM baselines in diverse areas while achieving favorable trade-offs in parameter count, runtime, and memory usage:

• 3D Human Mesh Recovery: Achieves state-of-the-art on 3DPW, MPI-INF-3DHP, outperforming previous HMR methods on both accuracy and temporal consistency (Chen et al., 29 Jan 2026).
• Multi-organ Segmentation: CS-MUNet delivers mDice = 86.16% (UW-Madison), 94.47% (WORD) with favorable efficiency compared to TransUNet/Swin-UNet (Zheng et al., 20 Mar 2026).
• Shadow Removal: D2-Mamba improves ALL-PSNR by up to 1.8 dB over ShadowMamba and 0.3 dB over HomoFormer, with only 9.39 M parameters (Li et al., 18 Aug 2025).
• Channel Estimation (OFDM): STA-Mamba yields MSE = 1.9e-3, BER = 1.8e-4 at 30 dB SNR using 0.35 M params (Channelformer: 0.93 M, HA02: 1.1 M), confirming parameter/memory efficiency (Luan et al., 23 Jan 2026).
• Bidirectional Time-Series (AnDi-2): Bi-Mamba achieves F1 = 0.91 for state segmentation, K MSLE = 0.05, $$h_t = \bar{A}_t h_{t-1} + \bar{B}_t x_t, \;\; y_t = C_t h_t + D x_t,$$2 MAE = 0.27, with excellent training stability (Lavaud et al., 2024).

These gains are consistently attributed to linear complexity in sequence length, adaptive long-range modeling, improved global/semantic context, and stable fusion/delineation of region-specific features.

6. Variants: Locally Bi-Directional and Dual-Input Extensions

While global bidirectional scans are the most common, variants exist:

• Locally Bi-directional Mamba (LBMamba) (Zhang et al., 19 Jun 2025): Embeds a register-resident, short-range backward scan within each thread's global forward selective scan. Full receptive field is recovered by stacking $$h_t = \bar{A}_t h_{t-1} + \bar{B}_t x_t, \;\; y_t = C_t h_t + D x_t,$$3 blocks, alternating scan directions. This approach reduces the doubled compute/memory overhead of naïve dual-scan implementations at negligible throughput cost.
• Dual-Input Semi-Supervised MambaX-Net (longitudinal MRI) (Yahathugoda et al., 20 Oct 2025): Fuses two time-adjacent scans using shared-weight encoders, a Mamba-enhanced Cross-Attention Module for feature temporal alignment, and a shape extractor for anatomical prior injection. Outperforms Transformer and SSM baselines on longitudinal prostate segmentation without human annotation, showing resilience to pseudo-label noise.

7. Limitations and Application-Specific Considerations

Limitations and caveats of Dual-Scan STA-Mamba schemes are noteworthy:

• HiSTM (Bettouche et al., 7 Aug 2025): Despite mentioning a "dual spatial encoder," does not implement an explicit dual-scan STA-Mamba; its hierarchy consists of single-pass 2D CNN and parallel per-pixel SSMs, with only temporal attention aggregation.
• Computational cost: Naive global dual-scan doubles FLOPs and memory bandwidth; register-local bi-directional scans (e.g., LBMamba) mitigate overheads (Zhang et al., 19 Jun 2025).
• Scan directionality vs. semantic alignment: The interpretability and efficacy of custom scanning (e.g., mask-aware, kinematic tree) can be highly domain-specific. Fusion and modulation must be tuned to avoid overfitting or underutilization of available cues (Li et al., 18 Aug 2025, Zheng et al., 20 Mar 2026).

Key references describing representative Dual-Scan STA-Mamba frameworks and their application-specific instantiations include:

• "Towards Geometry-Aware and Motion-Guided Video Human Mesh Recovery" (Chen et al., 29 Jan 2026)
• "MambaNet: Mamba-assisted Channel Estimation Neural Network With Attention Mechanism" (Luan et al., 23 Jan 2026)
• "CS-MUNet: A Channel-Spatial Dual-Stream Mamba Network for Multi-Organ Segmentation" (Zheng et al., 20 Mar 2026)
• "Bidirectional Mamba state-space model for anomalous diffusion" (Lavaud et al., 2024)
• "MambaTAD: When State-Space Models Meet Long-Range Temporal Action Detection" (Lu et al., 22 Nov 2025)
• "D2-Mamba: Dual-Scale Fusion and Dual-Path Scanning with SSMs for Shadow Removal" (Li et al., 18 Aug 2025)
• "ESTM: An Enhanced Dual-Branch Spectral-Temporal Mamba for Anomalous Sound Detection" (Ma et al., 2 Sep 2025)
• "LBMamba: Locally Bi-directional Mamba" (Zhang et al., 19 Jun 2025)
• "MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation" (Yahathugoda et al., 20 Oct 2025)

These works collectively establish the dual-scan state-space paradigm as an efficient and generalizable framework for long-range, context-aware neural sequence modeling across spatial, temporal, spectral, and semantic axes.