Hybrid CNN-Mamba Enhancement Network

• HCMEN is a hybrid architectural paradigm that combines CNN-based local feature extraction with state-space Mamba modules for efficient long-range dependency modeling.
• It employs a dual-path strategy with a convolution-dominant local pathway and a selective-scan global pathway that are fused through task-specific reconstruction techniques.
• Empirical evaluations across pan-sharpening, dehazing, speech enhancement, and medical segmentation highlight HCMEN’s practical trade-offs in efficiency and performance.

Hybrid CNN-Mamba Enhancement Network (HCMEN), used here as an Editor’s term, denotes a class of hybrid dense-prediction architectures that combine convolutional operators for local feature extraction, shallow alignment, or reconstruction with Mamba/state-space modules for efficient long-range dependency modeling. Across pan-sharpening and zero-shot multispectral super-resolution, image dehazing, cross-corpus speech enhancement, multimodal medical fusion, depth completion, and 2D/3D segmentation, the recurring decomposition is a convolution-dominant local pathway plus a selective-scan-based global pathway, followed by task-specific fusion and reconstruction (Wang et al., 17 Dec 2025, Wang et al., 1 Jul 2025, Kühne et al., 2 Oct 2025, Zhou et al., 5 Aug 2025, Xie et al., 27 May 2025). This suggests that HCMEN is better understood as an architectural paradigm than as a single standardized model name.

1. Conceptual scope and representative systems

The HCMEN pattern recurs in multiple forms. In some systems the hybridization is explicit in the main pipeline, as in MMMamba for pan-sharpening and zero-shot multispectral super-resolution, Laplace-Mamba for dehazing, ClinicalFMamba for multimodal neuroimaging fusion, and HTMNet for transparent and reflective object depth completion (Wang et al., 17 Dec 2025, Wang et al., 1 Jul 2025, Zhou et al., 5 Aug 2025, Xie et al., 27 May 2025). In others it appears as a latent-stage or stage-wise reallocation of roles, as in MedSegMamba, HyM-UNet, ACM-UNet, MambaVesselNet++, HC-Mamba, CM-UNet, CVMH-UNet, VCMamba, and HybridMamba (Cao et al., 2024, Chen et al., 22 Nov 2025, Huang et al., 30 May 2025, Xu et al., 26 Jul 2025, Xu, 2024, Liu et al., 2024, Cao et al., 2024, Munir et al., 4 Sep 2025, Wu et al., 18 Sep 2025).

Model	Domain	Hybrid pattern
MMMamba (Wang et al., 17 Dec 2025)	Pan-sharpening, zero-shot MS SR	Gated-conv dual encoder + Mamba fusion core + conv decoder
Laplace-Mamba (Wang et al., 1 Jul 2025)	Image dehazing	Low-frequency Mamba branch + high-frequency CNN branch
ClinicalFMamba (Zhou et al., 5 Aug 2025)	2D/3D medical fusion	DGCB CNN encoder + latent Mamba fusion + CMCA refinement
MedSegMamba (Cao et al., 2024)	3D brain segmentation	3D CNN encoder/decoder + SS3D Mamba bottleneck
HyM-UNet (Chen et al., 22 Nov 2025)	Medical segmentation	Shallow CNN stages + deep Visual Mamba + guided skip fusion
ACM-UNet (Huang et al., 30 May 2025)	Medical segmentation	Pretrained ResNet + VMamba via adapters + wavelet decoder
CM-UNet (Liu et al., 2024)	Remote-sensing segmentation	CNN encoder + attention-gated Mamba decoder
CVMH-UNet (Cao et al., 2024)	Remote-sensing segmentation	CVSSBlock global Mamba/local CNN + DCT skip fusion
HybridMamba (Wu et al., 18 Sep 2025)	3D medical segmentation	Global/local Mamba scans + FFT-gated spatial-frequency fusion
Mamba-CNN (Boukhari, 1 Sep 2025)	Facial beauty prediction	CNN backbone with simplified SSM-inspired gate

The family resemblance is architectural rather than nominal. Some papers use Mamba only at the bottleneck, some only in deep stages, some in parallel with a local branch, and some couple it to explicit frequency decomposition. A consistent theme is that convolutional inductive bias is retained where local structure is fragile, while selective scan is used where receptive-field expansion or multimodal aggregation becomes the primary bottleneck.

2. Architectural decomposition and division of labor

A central HCMEN principle is role specialization. CNN modules are assigned to local or shallow processing, while Mamba modules are assigned to global or latent processing. In MedSegMamba, the encoder is pure 3D CNN, the bottleneck is a sequence of nine VSS3D blocks built around SS3D selective scan, and the decoder is pure 3D CNN; Mamba is deliberately concentrated in the compressed latent space rather than distributed throughout the whole network (Cao et al., 2024). HyM-UNet makes the same separation hierarchically: stages 1–2 use residual convolution blocks to preserve high-frequency texture details, whereas stages 3–4 use Visual Mamba blocks to capture long-range semantic dependencies with linear complexity (Chen et al., 22 Nov 2025). VCMamba generalizes the same idea to a vision backbone with a convolutional stem and early convolutional stages, followed by late multi-directional Mamba blocks in stage 4 (Munir et al., 4 Sep 2025).

A second decomposition places convolution around a Mamba core. MMMamba starts from an upsampled multispectral image $I_{ms}$ and a panchromatic image $I_p$, extracts shallow modality-specific features with gated convolutional encoders,

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$

passes them through stacked MMMamba blocks, decodes only the MS branch, and reconstructs the target as

$I_{hms}=F_{ms}^{final}+I_{ms}.$

The design is hybrid in three places: CNN/gated-conv stems, depth-wise convolution inside each MMMamba block, and a convolutional decoder plus residual reconstruction head (Wang et al., 17 Dec 2025). Laplace-Mamba makes the division even more explicit by assigning low-frequency structure restoration to a Mamba branch and high-frequency detail enhancement to a CNN branch (Wang et al., 1 Jul 2025). ClinicalFMamba follows the same latent-fusion logic with a CNN hybrid feature encoder, a latent Mamba fusion module, cross-modal channel attention, and a lightweight convolutional decoder (Zhou et al., 5 Aug 2025).

At the operator level, several papers invoke the standard SSM formulation

$$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$

with zero-order-hold discretization used to motivate linear-time selective scan (Chen et al., 22 Nov 2025, Xu, 2024). In practice, however, the distinctive contribution of HCMEN-style systems is usually not a new low-level SSM equation but a new allocation of CNN and Mamba roles around that operator. This suggests that the decisive design choice is often not “whether to use Mamba,” but “where to use it, with what local preconditioning, and under what reconstruction pathway.”

3. Sequence construction, scan order, and token geometry

In HCMEN systems, sequence construction is a first-order design variable rather than a minor implementation detail. MMMamba states this most explicitly. Its Multimodal Interleaved State Space Model tokenizes both PAN and MS features into non-overlapping patches across four scan directions, interleaves corresponding PAN and MS patches so that cross-modal tokens are adjacent in sequence order, and then performs alternating local selective scans across paired windows (Wang et al., 17 Dec 2025). The practical importance is quantified by its WV2 ablation: replacing interleaving with sequential concatenation drops performance from $42.3120$ PSNR to $36.4702$, with corresponding degradation in SSIM, SAM, and ERGAS, which the paper interprets as evidence that token ordering is crucial for Mamba-based multimodal fusion (Wang et al., 17 Dec 2025).

The same issue appears in volumetric data. MedSegMamba’s SS3D module chooses one of six axis permutations, extracts eight sequences per orientation through rotations, unfolding, and reversal, and across all six orientation groups defines 48 unique traversal paths over the latent volume (Cao et al., 2024). HybridMamba formalizes a related dual strategy through SoMamba and LoMamba: $$\text{SoMamba}(x)=\text{Mamba}(x_f)+\text{Mamba}(x_r),$$

$$\text{LoMamba}(x)=\text{Mamba}(x_{lf})+\text{Mamba}(x_{lr})+\text{Mamba}(x_{ls}),$$

where slice-oriented forward/reverse traversal captures global context and local-window forward/reverse/across-slice traversal preserves neighborhood coherence (Wu et al., 18 Sep 2025). ClinicalFMamba extends scan design to 3D multimodal fusion by replacing naive 2D scanning with tri-plane scanning along axial, coronal, and sagittal planes; its ablation reports a $-7.055$ PSNR drop on BraTS 3D fusion when tri-plane scanning is replaced by conventional 2D-only scanning (Zhou et al., 5 Aug 2025).

A related 2D formulation appears in VCMamba, where the final stage uses four spatially continuous scan paths and direction-aware updates with learnable directional parameters $I_p$0, then sums directional outputs before projection (Munir et al., 4 Sep 2025). The recurring implication is that selective scan inherits the biases of sequence order. HCMEN design therefore depends not only on state-space capacity but on how 2D or 3D geometry is serialized: interleaving, plane-wise traversal, local-window flattening, and multi-orientation scanning all function as structural priors.

4. Fusion, gating, and frequency-aware enhancement

HCMEN systems typically do not fuse CNN and Mamba streams by naive addition alone. They use gates, attention, or frequency priors to regulate how local and global features interact. Laplace-Mamba is the clearest frequency-domain formulation. Its Laplace-Frequency Transform Module separates an image into low-frequency and high-frequency components; low-frequency features are fused with spatial features by the Multi-Domain Fusion Module,

$I_p$1

then processed by a low-frequency Mamba branch, while high-frequency features are refined by a CNN branch guided by the restored low-frequency output (Wang et al., 1 Jul 2025). This assigns global structure to Mamba and edge/detail restoration to CNNs on purpose.

ClinicalFMamba uses a different fusion strategy. Its Dilated Gated Convolution Block employs $I_p$2 and $I_p$3 convolutions with dilation rates $I_p$4, while Cross-Modal Channel Attention computes average-pooled and max-pooled channel descriptors from one modality to reweight the other, in both directions, before element-wise addition back to the latent features (Zhou et al., 5 Aug 2025). CM-UNet’s CSMamba block similarly uses channel and spatial attention as the gate activation condition of vanilla Mamba: $I_p$5

$I_p$6

$I_p$7

so the Mamba branch is explicitly modulated by local saliency priors (Liu et al., 2024).

Skip-path refinement is another recurrent enhancement mechanism. HyM-UNet’s Mamba-Guided Fusion Skip Connection computes

$I_p$8

$I_p$9

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$0

using deep semantic decoder features to suppress noisy shallow activations while preserving boundaries through residual enhancement (Chen et al., 22 Nov 2025). CVMH-UNet’s MFMSBlock applies an analogous soft weighting between encoder and decoder features,

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$1

where $$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$2 incorporates DCT-based multi-frequency channel attention and a local point-wise branch (Cao et al., 2024). HybridMamba pushes frequency fusion further by constructing

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$3

and blending it with spatial features through

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$4

with learnable frequency thresholds $$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$5 and $$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$6 at initialization (Wu et al., 18 Sep 2025).

Taken together, these mechanisms suggest that HCMEN is not merely “CNN front-end plus Mamba backbone.” A defining characteristic is learned arbitration between local and global evidence, often mediated by channel attention, spatial attention, spectral decomposition, or frequency-domain priors.

5. Objectives, implementation patterns, and efficiency

Training objectives in HCMEN systems are usually simple at the pixel or task level, then augmented by structure-aware or frequency-aware terms when the domain requires them. MMMamba uses

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$7

for reduced-resolution pan-sharpening under the Wald protocol (Wang et al., 17 Dec 2025). Laplace-Mamba combines a spatial reconstruction loss with a frequency loss,

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$8

so the decomposition used by the architecture is also supervised (Wang et al., 1 Jul 2025). ClinicalFMamba adopts a hybrid fusion objective,

$$F_{ms}=E_{\varphi}^{ms}(I_{ms}),\qquad F_p=E_{\varphi}^{p}(I_p),$$9

with $I_{hms}=F_{ms}^{final}+I_{ms}.$0, $I_{hms}=F_{ms}^{final}+I_{ms}.$1, and $I_{hms}=F_{ms}^{final}+I_{ms}.$2, and extends gradient and SSIM terms to 3D variants for volumetric fusion (Zhou et al., 5 Aug 2025). Segmentation-oriented members of the family retain task-specific supervision, such as Dice plus BCE plus edge loss in HyM-UNet or Dice plus cross-entropy in ACM-UNet (Chen et al., 22 Nov 2025, Huang et al., 30 May 2025).

Efficiency claims are a major part of the HCMEN literature, but they are nuanced rather than uniform. MMMamba reports $I_{hms}=F_{ms}^{final}+I_{ms}.$3 G FLOPs and $I_{hms}=F_{ms}^{final}+I_{ms}.$4 M parameters under the reported input size (Wang et al., 17 Dec 2025). Laplace-Mamba reports $I_{hms}=F_{ms}^{final}+I_{ms}.$5 GFLOPs, $I_{hms}=F_{ms}^{final}+I_{ms}.$6 s runtime on $I_{hms}=F_{ms}^{final}+I_{ms}.$7 images, and $I_{hms}=F_{ms}^{final}+I_{ms}.$8 M parameters for its best configuration (Wang et al., 1 Jul 2025). ClinicalFMamba reports $I_{hms}=F_{ms}^{final}+I_{ms}.$9 M parameters and $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$0 s per $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$1 image pair in 2D, and $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$2 M parameters with $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$3 s per $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$4 volume pair in 3D (Zhou et al., 5 Aug 2025). ACM-UNet reports $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$5 M parameters and $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$6 G FLOPs for $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$7 input, while VCMamba-B reports $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$8 M parameters and $$h'(t)=\mathbf{A}h(t)+\mathbf{B}x(t),\qquad y(t)=\mathbf{C}h(t),$$9 GMACs on ImageNet-1K (Huang et al., 30 May 2025, Munir et al., 4 Sep 2025). MedSegMamba emphasizes a different systems trade-off: about 20% fewer parameters than SegMambaBot, similar inference memory, but slower execution because richer SS3D traversal increases scan cost (Cao et al., 2024).

This suggests that the standard “Mamba is linear” argument is necessary but insufficient. In HCMEN practice, end-to-end cost depends on scan multiplicity, local preconditioning layers, frequency transforms, multibranch fusion, and decoder topology, not on the selective scan operator alone.

6. Empirical evidence, misconceptions, and limitations

Empirically, HCMEN-style systems are supported across a wide task range. MMMamba reports the best reduced-resolution pan-sharpening results across most metrics on WV2, GF2, and WV3, and also performs zero-shot multispectral super-resolution by dropping the PAN modality at inference time (Wang et al., 17 Dec 2025). Laplace-Mamba reports $42.3120$0 PSNR and $42.3120$1 SSIM on Haze4K, $42.3120$2 on LMHaze, and $42.3120$3 on O-Haze, together with favorable efficiency (Wang et al., 1 Jul 2025). RWSA-MambaUNet shows that a hybrid Mamba-attention U-Net can dominate cross-corpus speech enhancement on out-of-domain DNS 2020 and EARS-WHAM$42.3120$4 benchmarks while using markedly fewer FLOPs than the strongest baselines (Kühne et al., 2 Oct 2025). ClinicalFMamba not only improves fusion metrics on MRI-CT, MRI-SPECT, and BraTS, but also improves downstream brain-tumor grading AUC and F1 on fused inputs (Zhou et al., 5 Aug 2025). HyM-UNet, ACM-UNet, MambaVesselNet++, and HybridMamba show that the same hybrid logic transfers to 2D and 3D medical segmentation, especially when boundary fidelity and multiscale structure matter (Chen et al., 22 Nov 2025, Huang et al., 30 May 2025, Xu et al., 26 Jul 2025, Wu et al., 18 Sep 2025).

Several common misconceptions are corrected by these papers. First, not every “Mamba-CNN” is a full selective-scan architecture. Mamba-CNN explicitly states that its MambaBlock is a simplified SSM-inspired gating mechanism built from depthwise convolution and sigmoid gating, and that it does not implement the full selective scan parameterization of original Mamba (Boukhari, 1 Sep 2025). Second, more Mamba is not automatically better. ACM-UNet reports its best Synapse result with 2 VSS blocks rather than 4 or 8, Laplace-Mamba reports performance degradation when LSRB/HDEB depth is increased beyond its best configuration, and VCMamba’s ablation shows that interleaving Mamba and convolutional FFN is superior to a more naive stage replacement strategy (Huang et al., 30 May 2025, Wang et al., 1 Jul 2025, Munir et al., 4 Sep 2025). Third, some narratives overstate their own tables. MMMamba’s text claims to “consistently” outperform others on full-resolution GF2, but its table shows mixed best values across $42.3120$5, $42.3120$6, and QNR (Wang et al., 17 Dec 2025).

Reproducibility remains uneven. Many papers omit exact state dimensions, scan merge rules, patch or crop sizes, augmentation policies, optimizer schedules, or decoder details. MMMamba does not expose low-level Mamba equations or several training details in the provided method section, MedSegMamba omits exact channel widths and upsampling operator details, ClinicalFMamba omits optimizer and LR settings for the fusion network, and several segmentation papers leave adapter or decoder internals underspecified (Wang et al., 17 Dec 2025, Cao et al., 2024, Zhou et al., 5 Aug 2025, Huang et al., 30 May 2025). A further limitation is task specificity: interleaving schemes assume spatial correspondence in multimodal fusion, tri-plane scanning assumes anatomically meaningful orthogonal planes, and frequency-gated modules may be sensitive to degradation type. The broader implication is that HCMEN is a strong design language, but not yet a fully standardized recipe.

The most stable conclusion is architectural. HCMEN is most effective when CNN components preserve locality, frequency selectivity, and reconstruction fidelity; Mamba components model long-range dependencies where token count is manageable; and fusion is treated as a learned, content-dependent operation rather than a fixed concatenation. In that sense, the family’s defining lesson is not merely “combine CNN and Mamba,” but “hybridize them where their inductive biases are strongest, and make token order, frequency structure, and skip refinement explicit design variables.”