Mamba+CNN Hybrid Architecture

• Mamba+CNN Hybrids are neural architectures that combine CNN modules for precise local feature extraction with Mamba state-space models for efficient global context aggregation.
• They employ diverse fusion strategies—such as encoder-decoder splits, alternating mixers, and adapter-mediated integration—to seamlessly merge local details with long-range dependencies.
• Empirical evidence in applications like semantic segmentation, object detection, and medical imaging shows these hybrids deliver improved metrics and reduced computational costs compared to traditional CNN-only designs.

A Mamba+CNN Hybrid designates neural architectures that combine convolutional neural networks (CNNs) for local feature extraction with Mamba state-space models (SSMs) for efficient long-range dependency modeling. This hybridization addresses the limitations of CNNs—strong locality bias but limited receptive field—and Transformers—high computational complexity in attention—by leveraging Mamba’s linear-time global sequence modeling. Mamba+CNN Hybrids have demonstrated superior accuracy–efficiency tradeoffs and enhanced feature representation across diverse domains, including semantic segmentation, object detection, medical imaging, and multimodal learning.

1. Architectural Principles and Hybridization Strategies

Most Mamba+CNN Hybrids implement an explicit separation of roles: early network stages use depthwise or pointwise convolutions to extract spatially local, high-frequency features, while deeper stages or dedicated modules invoke Mamba (or VMamba) SSMs for global/contextual aggregation. Architectures vary in the fusion mechanism and locus of hybridization, including encoder–decoder splits, token-mixer alternation within a block, and parallel residual-branch strategies.

Typical patterns are:

• UNet-style encoder-decoder: Convolutional encoder (e.g., ResNet-18/ResNet-50) is paired with a Mamba-based decoder (e.g., CSMamba or VSS blocks); skip connections may include Mamba-guided fusion gates (Liu et al., 17 May 2024, Chen et al., 22 Nov 2025).
• Alternating or bottlenecked mixing: Interleaving convolutional token mixers (e.g., Inception-like, banded, or dual-attention) with bottleneck Mamba modules acting as global mixers (Wang et al., 10 Jun 2025).
• Adapter-mediated integration: Lightweight channel adapters bridge architectural/semantic gaps between CNN and Mamba pre-trained backbones (Huang et al., 30 May 2025).
• Capsule/capsuled feature fusions: Capsule-style fusion modules (e.g., FFM with 2D-SSM) knit together CNN features from multiple scales using Mamba linear scan (Du et al., 10 Jun 2025).
• Multi-view and multi-branch: Mamba is used alongside CNN (and occasionally Transformer) branches, and co-attention gates aggregate cross-branch information (e.g., MambaCAFU (Bui et al., 4 Oct 2025), RemoteDet-Mamba (Ren et al., 17 Oct 2024)).

2. Core Components and Mathematical Formulation

Table: Representative hybrid modules

Module/Class	CNN Role	Mamba Role	Fusion Mechanism
CSMamba block (Liu et al., 17 May 2024)	Local feature extraction	2D-SSM for long-range	Gated Hadamard, channel+spatial attention
Bottleneck Mamba (Wang et al., 10 Jun 2025)	Convolutional mixer	SS2D scan	Residual, depthwise+MLP
Hi-Encoder (Xu et al., 26 Jul 2025)	Texture-aware layers	Vision Mamba	Sequential: conv→SSM, skip-addition
Adapter (Huang et al., 30 May 2025)	Pretrained ResNet	Pretrained VMamba	Two-layer linear channel projections
MambaBlock (Boukhari, 1 Sep 2025)	Depthwise/inv-residual	SSM-inspired gating	Parallel, feature modulation

Hybrid modules frequently implement Mamba as a linear 2D or 3D selective scan over feature sequences $x_t$, with updates: $$h_t = A h_{t-1} + B x_t, \;\;\; y_t = C h_t + D x_t$$ where the mapping from local features $x_t$ to global context $h_t$ is learned, and directional variants (e.g., quad/triplane scans) ensure full spatial coverage with linear $\mathcal{O}(N)$ cost (Zhou et al., 5 Aug 2025, Cao et al., 12 Sep 2024, Munir et al., 4 Sep 2025). Fusion mechanisms include skip-residual addition, Hadamard gating, channel/spatial attention, or explicit co-attention.

3. Domain-Specific Instantiations

Remote Sensing: CM-UNet employs ResNet-18 encoding and a CSMamba-based decoder; a Multi-Scale Attention Aggregation (MSAA) module refines encoder-side features. The CSMamba block introduces joint channel–spatial gating as a condition for vanilla Mamba to enhance global-local feature interaction. On ISPRS Potsdam, it achieves mIoU=87.21%, mF1=93.05%, and OA=91.86% (Liu et al., 17 May 2024).

Medical Image Segmentation:

• HyM-UNet alternates CNN and Visual Mamba blocks in a hierarchical encoder, using Mamba-Guided Fusion Skip Connections (MGF-Skip) to suppress background noise in shallow features, resulting in 88.97% Dice and 81.82% IoU on ISIC 2018 (Chen et al., 22 Nov 2025).
• ACM-UNet stages ResNet-50 encoding with VMamba-based state-space blocks (with adapters), and a multi-scale wavelet transform (MSWT) decoder, reporting 85.12% Dice on Synapse and 92.29% on ACDC (Huang et al., 30 May 2025).
• MambaVesselNet++ utilizes texture-aware convolutional encoding followed by stacked Vision Mamba layers, and a bifocal fusion decoder. This design achieves 0.953 Dice on PH2 and 0.870 on 3D IXI vascular segmentation (Xu et al., 26 Jul 2025).
• ECMNet employs pure CNN encoder–decoder blocks, augmented with an efficient Mamba-based Feature Fusion Module (FFM) in the bottleneck, and achieves 73.6% mIoU (CamVid) with only 0.87M parameters (Du et al., 10 Jun 2025).

Object Detection:

• MambaNeXt-YOLO integrates depthwise separable ConvNeXt blocks (local) and Mamba SSM global modules via an adaptive ResGate, both in its backbone and pyramid neck, yielding 66.6% mAP at 31.9 FPS on Pascal VOC with 7.1M parameters (Lei et al., 4 Jun 2025).

Multimodal and Multitask Learning:

• HTMNet applies a CNN+Transformer encoder and a Transformer-Mamba bottleneck fusion for multimodal RGB-D completion, demonstrating state-of-the-art error rates for transparent/reflective depth estimation (Xie et al., 27 May 2025).
• RemoteDet-Mamba fuses the outputs of parallel Siamese CNN branches (for RGB and TIR) using a quad-directional selective scan Cross-modal Fusion Mamba (CFM), reporting [email protected]=0.818 on DroneVehicle (Ren et al., 17 Oct 2024).
• HybridMamba alternates slice-oriented and local-adaptive Mamba mechanisms with frequency-gated CNNs for 3D segmentation, outperforming pure Mamba in both Dice coefficient and HD95 for BraTS and lung cancer datasets (Wu et al., 18 Sep 2025).

4. Computational Efficiency and Scaling

Mamba+CNN Hybrids consistently demonstrate favorable accuracy–efficiency tradeoffs:

• Mamba modules provide linear complexity in sequence or path length, compared with quadratic scaling of transformer attention (Munir et al., 4 Sep 2025, Liu et al., 17 May 2024).
• Parameter and FLOP counts are modest; e.g., ECMNet: 0.87M params/8.3G FLOPs for 73.6% mIoU, ACM-UNet: 17.93G FLOPs/16.48M params for 85.12% Dice, MambaNeXt-YOLO: 7.1M params/22.4G FLOPs (Du et al., 10 Jun 2025, Huang et al., 30 May 2025, Lei et al., 4 Jun 2025).
• In large 3D-volume settings (e.g., MedSegMamba, HybridMamba), Mamba blocks—using patch-based 3D SS3D traversals with multiple axis- or view-oriented selective scan sequences—avoid the memory explosion typical of transformer attention, maintaining tractable GPU usage (Cao et al., 12 Sep 2024, Wu et al., 18 Sep 2025).
• Fast inference: HyM-UNet achieves 23ms per 256×256 image with ≈8.2M parameters (Chen et al., 22 Nov 2025); MambaCAFU variants operate within the 30–70M parameter range but exceed pure transformer benchmarks (Bui et al., 4 Oct 2025).

5. Empirical Gains and Ablation Evidence

Ablation studies in nearly all cited works support the complementarity of CNNs for local detail and Mamba for global structure:

• In HyM-UNet, adding Mamba yields +1.02% IoU; adding MGF-Skip a further +1.48% IoU. In ACM-UNet, adapters + MSWT + VSS each provide measurable Dice/HD95 gains (Chen et al., 22 Nov 2025, Huang et al., 30 May 2025).
• In ECMNet, Mamba-FFM alone yields +0.83% mIoU over baseline; skip connections with MSAU attention add another +2%; full hybrid provides +3.7% vs. CNN-only (Du et al., 10 Jun 2025).
• In object detection and recognition backbones (e.g., MambaNeXt-YOLO, VCMamba), alternating or integrating CNN and Mamba blocks either match or exceed larger Transformer or SSM-only designs at a fraction of the cost (Lei et al., 4 Jun 2025, Munir et al., 4 Sep 2025).
• On mammography, brain MRI, ultrasound, and vascular segmentation, hybrid models regularly set new benchmarks or cut resource usage by over 10× with no drop in segmentation fidelity (Bui et al., 4 Oct 2025, Han et al., 28 Sep 2024, Li et al., 1 Jan 2025, Xu et al., 26 Jul 2025).

6. Advantages, Limitations, and Adaptation Guidelines

Advantages:

• Linear-time long-range context propagation, enabling high-resolution or volumetric modeling impractical for transformers.
• Explicit local–global division of labor: CNNs maintain spatial/edge fidelity; Mamba SSMs inject image-wide or volumewide structure.
• Parameter and computational efficiency relative to both transformer-heavy and deep CNN-only designs.
• Robustness in challenging settings: noisy modalities, missing modalities, sparse annotations, or multimodal fusion (e.g., MRI-CT, RGB-D, multi-sensor).

Limitations:

• Slight increase in system complexity (hybrid block wiring, SSM tuning).
• Domain-specific adaptations: e.g., 2D vs. 3D scan orders, axial views, gating in anisotropic medical volumes may require specialized engineering (Chen et al., 22 Nov 2025).
• Some pipelines (e.g., evidence-guided consistency in MambaEviScrib) introduce extra supervision and calibration complexity (Han et al., 28 Sep 2024).

Adaptation Guidance:

• For high-resolution 2D/3D inputs: use CNNs for early, fine-grained spatial encoding followed by Mamba at lower spatial resolutions.
• Decoder/skip fusion: employ attention-gated or co-attention-based channel fusion for robust boundary and region prediction.
• Mamba block depth and scan order should be tuned to available compute and volumetric structure; hybridization locus can be adapted between encoder, decoder, or parallel branches depending on task.
• Maintain plug-and-play channel adapters when integrating off-the-shelf CNNs and VMamba or SSM blocks (Huang et al., 30 May 2025).

7. Impact and Outlook

Mamba+CNN Hybrids have established themselves as a versatile backbone for image segmentation, detection, fusion, and regression in computationally demanding domains—most notably in medical imaging and remote sensing. The linear scaling of Mamba blocks enables practical deployment on high-resolution and 3D datasets. Empirical benchmarks demonstrate consistent improvement over both CNN-only and Transformer-based methods at comparable or superior efficiency, often with lower parameter counts and inference latency.

The field is trending toward more sophisticated multi-branch and attention-modulated hybrids, combinatorial use of Mamba with Transformers and channel-coattention gates, and adaptive module selection based on data structure (e.g., co-attention in medical imaging, capsule-based FFMs, and gated fusion in cross-modal architectures). Ablation evidence from a wide range of benchmarks substantiates the “best of both worlds” advantage predicted by the hybrid paradigm (Liu et al., 17 May 2024, Huang et al., 30 May 2025, Chen et al., 22 Nov 2025). Continued research is anticipated in extending these principles to other domains requiring high fidelity, efficient, and globally consistent representations.