SRMA-Mamba: Spatial Reverse Mamba Networks

• SRMA-Mamba is a deep learning framework that employs selective state-space models with spatial and temporal attention to efficiently capture long-range dependencies.
• It utilizes domain-specific scanning and aggregation techniques—like SABMamba and SRMA blocks—to achieve superior results in 3D segmentation, MRI super-resolution, and point cloud upsampling.
• The architecture advances performance by replacing quadratic self-attention with linear-time operations, yielding improved metrics such as Dice coefficients, PSNR, and IoU in diverse applications.

SRMA-Mamba ("Spatial Reverse Mamba Attention Network") is a designation used for multiple recent architectures built on the Mamba selective state-space model (SSM) paradigm, with applications spanning 3D medical image segmentation, super-resolution in MRI and hyperspectral imagery, surgical video analysis, and point cloud upsampling. The SRMA-Mamba models leverage the linear-time global modeling capacity of SSMs while integrating spatially and/or temporally structured mechanisms tailored to domain-specific tasks such as anatomical context aggregation, pixel-wise sequential interaction, and multi-directional feature encoding.

1. Foundational Principles: Mamba State-Space Models and Selective Context Encoding

The core technical innovation across SRMA-Mamba variants is the deployment of Mamba-style SSMs, which generalize discretized linear time-invariant dynamical systems for deep learning. Each Mamba block maintains a hidden state updated recursively via learned matrices $\overline{A},\overline{B}$, consuming an input sequence $\{x_t\}$ and producing outputs $y_t = C\,h_t + D\,x_t$, with

$h_{t} = \overline{A}\,h_{t-1} + \overline{B}\,x_t.$

Mamba replaces quadratic-complexity self-attention with $O(L)$-time sequence modeling and supports input-dependent parameterization (e.g., step-size $\Delta_t$) via in-block MLPs, enabling efficient long-range modeling in high-dimensional temporal, spatial, or spatiotemporal domains (Zeng et al., 17 Aug 2025, Safari et al., 22 Dec 2025).

Domain-specific SRMA-Mamba designs typically augment SSMs with structured scanning and aggregation:

• Spatial Context: Orthogonal-axis decomposition, local 3D partitioning, or axis-specific scanning for MRI, CT, and hyperspectral volumes (Zeng et al., 17 Aug 2025, Chen et al., 30 Jan 2025).
• Temporal Context: Bidirectional recurrence or causal masking in video and sequential analysis (Cao et al., 2024).
• Pixel/Channel Recursion: Layer-wise feature recycling for SR (He et al., 2024).

2. SRMA-Mamba in Volumetric Medical Image Segmentation

The canonical SRMA-Mamba network for 3D liver MRI segmentation (Zeng et al., 17 Aug 2025) employs a U-Net style encoder–decoder backbone. Each encoder stage consists of a Gated Spatial Convolution (GSC) block followed by the Spatial Anatomy-Based Mamba (SABMamba) block:

• SABMamba Block: Features are decomposed into three orthogonal 2D anatomical planes—sagittal, coronal, and axial. Within each plane, CrossScan and CrossMerge mechanisms apply directional Mamba scans, capturing long-range dependencies along each anatomical axis with linear time complexity. The outputs are aggregated to form a volumetric context representation, crucial for segmenting structures with high anatomical variability.
• Spatial Reverse Mamba Attention (SRMA): Decoder modules iteratively refine the segmentation, computing reverse attention weights by "erasing" confident predictions ($A_i = 1-\sigma(S_{i+1})$) and re-focusing on difficult boundaries and small nodules through SABMamba-mediated feature refinement layers.

This composite architecture outperforms comparable SSM and attention-based networks, delivering Dice coefficients of 92.95% (T1W) and 86.25% (T2W), and substantially reducing boundary errors such as the 95th percentile Hausdorff distance (HD95) from 26.97 mm to 17.52 mm compared to SegMamba. Ablations reveal that omitting global spatial context (SABMamba) or boundary refinement (SRMA) results in dramatic performance drops (Zeng et al., 17 Aug 2025).

3. Vision SRMA-Mamba for Super-Resolution in MRI and Hyperspectral Imaging

SRMA-Mamba architectures targeting 2D or 3D super-resolution tasks instantiate advanced SSM-based feature extraction with domain-aligned scanning strategies.

• MRI Super-Resolution: The Vision Mamba SR framework (Safari et al., 22 Dec 2025) utilizes a hybrid selective scanning module, extracting horizontal, vertical, and diagonal patch-wise sequences. Each is processed via a multi-head selective SSM (MHSSM), and the outputs are fused with a channel-gated MLP, capturing both local and global anatomical detail. A minimal parameter count (0.9M) and compute budget (57 GFLOPs) deliver state-of-the-art results (SSIM = 0.951, PSNR = 26.90 dB) on 7T brain MRI, outperforming Res-SRDiff, SwinIR, and MambaIR in both performance and efficiency.
• Hyperspectral SR (HSRMamba): Sequence modeling across the entire $H \times W \times B$ cube is enhanced by local spatial–spectral partitioning (LSSP), combining local block-wise causal SSM passes (with masked convolutional kernels) with global spectral reordering (GSRM) that permutes bands by spectral similarity before BSSM processing. This dual mechanism alleviates "local forgetting" and order sensitivity inherent in vanilla vectorized SSMs, achieving superior quantitative (PSNR 40.28 dB, SSIM 0.9441) and qualitative reconstruction compared to group CNNs and transformer variants (Chen et al., 30 Jan 2025).

4. Specialized SRMA-Mamba Variations: Pixel-Wise SR, Surgical Video, and Point Cloud Applications

The SRMA-Mamba paradigm extends to other domains where long-range dependency modeling is critical.

• Pixel-Wise Image SR: The MPSI network (He et al., 2024)—also referenced as SRMA-Mamba—introduces a Channel-Mamba Block (CMB) with bi-directional SSMs (forward/reverse) and a Mamba Channel Recursion Module (MCRM) for layerwise feature aggregation. These modules permit explicit modeling of pixel-pairwise relationships and multi-span channel context, outperforming lightweight and baseline transformer models on Urban100 and Manga109 ($+0.07$ dB PSNR gains over EDSR and SwinIR).
• Surgical Video Analysis: In surgical phase recognition, SR-Mamba (Cao et al., 2024) applies a bidirectional Mamba decoder atop spatial frame embeddings, enabling sequence lengths up to 2048. This configuration achieves SOTA recognition performance (92.6% accuracy, 81.5% Jaccard index on Cholec80), with architectural ablations confirming benefits for long context modeling and anticipation loss.
• LiDAR Point Cloud SR: SRMamba (Chen et al., 15 May 2025) projects sparse 3D point clouds into 2D range images, compensates for scan artifacts, and reconstructs HR depth maps using Visual State Space Models with multi-directional SSM scanning within an asymmetric U-Net encoder-decoder, yielding a $9.5$\% IoU improvement and substantial CD reduction versus TULIP, Swin-IR, and Cas-ViT on KITTI-360.

5. Implementation and Ablation Findings

All SRMA-Mamba variants employ deep learning protocols matched to their input structure and domain priors:

• Encoder depth and module stacking (e.g., 4-stage, 18-layer, or group-based setups) are optimized via grid search or ablation on benchmark splits (Zeng et al., 17 Aug 2025, He et al., 2024, Chen et al., 30 Jan 2025).
• Training regimes involve task-appropriate losses: Dice + cross-entropy for segmentation, $\{x_t\}$0 plus perceptual/LPIPS for SR, and custom regularizers (e.g., spectral norm penalties) for SSM stability.
• Key module ablation studies consistently show that the explicit SSM-based components (SABMamba, SRMA, CMB, MCRM, MHSSM) contribute absolute performance gains of 0.1–12%, context-dependent, while hybrid scanning and layerwise recursion further boost stability and fidelity.
• Data augmentation, adaptive learning rate, and input-specific preprocessing are standard across implementations.

6. Robustness and Limitations of SRMA-Mamba Models

Despite their modeling power, SRMA-Mamba models inherited vulnerabilities highlighted in (Latibari et al., 16 Feb 2026):

• Under white-box adversarial attacks (e.g., 20-step PGD, $\{x_t\}$1), robust accuracy collapses (e.g., 93.3%$\{x_t\}$20% on PneumoniaMNIST).
• PatchDrop occlusion, Gaussian noise, defocus blur, and hardware-induced bit flips (even at single-bit per critical module) cause sharp, module-localized performance drops.
• Early-stage (Patch_embed, first SSM) sensitivity is acute; failure propagates to downstream outputs.

Recommended defenses include adversarial and corruption training, spectral regularization, error correction coding, module-level redundancy, and run-time anomaly detection.

7. Comparative Performance and Impact

SRMA-Mamba and its derived architectures have demonstrated new state-of-the-art across several modalities and tasks: large-scale 3D medical segmentation, medical image SR, pixel-level image enhancement, surgical workflow analysis, and geometric upsampling in point clouds.

Key results include:

Application	SOTA Metric	SRMA-Mamba Value	Reference
3D Liver Segmentation	Dice (T1W)	92.95%	(Zeng et al., 17 Aug 2025)
MRI Super-Resolution (7T)	SSIM / PSNR	0.951 / 26.90 dB	(Safari et al., 22 Dec 2025)
Hyperspectral SR (Chikusei)	PSNR / SSIM	40.28 / 0.9441	(Chen et al., 30 Jan 2025)
Pixel-wise SR (Urban100)	PSNR	26.71 dB	(He et al., 2024)
Surgical Phase Recognition	Accuracy (Cholec80)	92.6%	(Cao et al., 2024)
LiDAR Super-Resolution	IoU (KITTI-360)	0.4548	(Chen et al., 15 May 2025)

These results display consistent, significant improvements over transformer, CNN, and diffusion-based baselines in both computational efficiency and accuracy, supporting the broad impact of the SRMA-Mamba family.

8. Limitations and Future Directions

Open challenges for SRMA-Mamba include:

• Overfitting at increased SSM depth, indicating the need for adaptive or sparsity-constrained depth scheduling (Chen et al., 30 Jan 2025).
• Structural limitations of fixed spatial partitioning or windowing, motivating dynamic or content-adaptive scanning strategies.
• Need for run-time and hardware safety mechanisms before clinical or real-time deployment due to discovered fragility to adversarial, occlusion, and bit-flip faults (Latibari et al., 16 Feb 2026).
• Incorporation of joint spatial and spectral patch permutations, compressive modeling, and quantization for resource-limited applications.

SRMA-Mamba models thus represent a rapidly advancing, SSM-driven design paradigm, yielding high-performance solutions for complex visual reasoning tasks but requiring further robustness work for safety-critical settings.