Mamba-CNN Hybrid Architectures

• Mamba-CNN is a hybrid architecture that combines CNNs for extracting local patterns with Mamba-inspired selective state space models for global context aggregation.
• It employs diverse integration patterns, such as plug-and-play modules and parallel branches, to optimize performance in tasks like semantic segmentation and object detection.
• The approach has demonstrated measurable improvements in domains such as facial beauty prediction, medical imaging, and biosignal analysis, validating its design choices.

Mamba-CNN denotes a family of hybrid architectures that combine convolutional neural networks (CNNs) with Mamba-based or Mamba-inspired selective state space mechanisms in order to couple local inductive bias with efficient long-range dependency modeling. In the computer-vision survey literature, this hybridization is motivated by the complementary limitations of the component families: CNNs are effective at extracting local spatial features but have difficulty capturing long-range dependencies without substantial architectural depth, whereas Vision Transformers model global relationships at quadratic cost; Mamba addresses the latter issue through Selective Structured State Space Models with linear computational complexity (Rahman et al., 2024). The term also appears as the proper name of a facial beauty prediction architecture that inserts a lightweight, Mamba-inspired State Space Model gating mechanism into a hierarchical convolutional backbone (Boukhari, 1 Sep 2025).

1. Conceptual basis

The core rationale of Mamba-CNN design is functional complementarity. Across the surveyed literature, CNN components are assigned local pattern extraction—edges, textures, short-range temporal motifs, or fine anatomical detail—while Mamba or visual state space components are assigned global or long-range propagation over serialized image, volume, or sequence representations (Rahman et al., 2024). This division of labor recurs in image classification, semantic segmentation, object detection, multimodal sentiment analysis, EEG decoding, ECG analysis, and biomedical 3D segmentation.

At the state-space level, several papers ground Mamba in the discrete recurrence

$h_k = A h_{k-1} + B x_k,\qquad y_k = C h_k,$

and emphasize that Mamba’s distinctive property is input-dependent selectivity. In the attention-oriented interpretation developed for crack segmentation, the parameters are made input dependent as

$$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$

which is presented as the source of “parameter selectivity” analogous to attention-like adaptation (He et al., 2024). This explains why hybrid designs often treat Mamba not as a wholesale replacement for CNNs, but as a module for context aggregation, sequence propagation, or gated modulation layered on top of convolutional feature extraction.

The same literature also frames the complexity trade-off explicitly. CNNs are local; Transformer self-attention is quadratic in input size; Mamba-based selective scan is linear, which makes it attractive in high-resolution vision and long-sequence biosignal settings (Rahman et al., 2024). This suggests that the Mamba-CNN paradigm is less a single architecture than a recurrent systems pattern for combining local and global operators under stricter memory and throughput budgets.

2. Integration patterns

The family is architecturally heterogeneous, but the reported designs cluster around a small number of recurring integration patterns: plug-and-play replacement of context heads, serial local-to-global blocks, parallel local/global branches, asymmetric encoder-decoder hybrids, and training-only collaborative dual branches.

Model	Domain	Integration pattern
"Generalizing Geometry-Guided Mamba as a Plug-and-Play Context Module for CNN-based Semantic Segmentation" (Chan et al., 7 Jun 2026)	Semantic segmentation	Replaces ASPP, PPM, or attention heads while keeping the backbone unchanged
"CM-UNet: Hybrid CNN-Mamba UNet for Remote Sensing Image Semantic Segmentation" (Liu et al., 2024)	Remote sensing segmentation	CNN encoder, MSAA bridge, CSMamba decoder
"MambaMIC: An Efficient Baseline for Microscopic Image Classification with State Space Models" (Zou et al., 2024)	Microscopic image classification	Channel-split local Conv branch and global SSM branch fused by MIFA
"MambaNeXt-YOLO: A Hybrid State Space Model for Real-time Object Detection" (Lei et al., 4 Jun 2025)	Object detection	ConvNeXt local branch, Mamba global branch, ResGate fusion, MAFPN
"SWIM: Short-Window CNN Integrated with Mamba for EEG-Based Auditory Spatial Attention Decoding" (Zhang et al., 2024)	EEG decoding	Short-window CNN feature extractor followed by Mamba sequence model
"Mamba-CNN: A Hybrid Architecture for Efficient and Accurate Facial Beauty Prediction" (Boukhari, 1 Sep 2025)	Facial beauty prediction	Hierarchical CNN backbone with Mamba-inspired gating and multi-scale pooling

These patterns are not interchangeable. Plug-and-play modules preserve the baseline backbone and decoder so that changes can be attributed to context aggregation alone, as in geometry-guided semantic segmentation (Chan et al., 7 Jun 2026). Serial designs place CNN before Mamba so that local structure is encoded prior to selective scan, as in multimodal sentiment analysis and ECG classification (Li et al., 31 Jul 2025, Jiang et al., 14 Oct 2025). Parallel designs split channels or branches so that convolutional and state-space pathways operate concurrently before fusion, as in microscopic image classification and lightweight visual backbones (Zou et al., 2024, He et al., 2024). Training-only dual-branch schemes use Mamba as a collaborator rather than as an inference-time component, as in scribble-supervised ultrasound segmentation, where only the U-Net branch is retained at test time (Han et al., 2024).

3. Dense prediction and detection

Semantic segmentation is the most extensive application area for Mamba-CNN hybrids. A particularly controlled study treats Geometry-Guided Mamba as a drop-in replacement for standard context heads in six CNN segmentation models—DeepLabV3+, DANet, CCNet, PSPNet, PSANet, and OCRNet—while keeping the ResNet-101 backbone, output stride 8, and training settings unchanged (Chan et al., 7 Jun 2026). The module predicts a centripetal potential map $V$, a directional flow field $\Phi$, and a coarse morphological boundary map $D_c$; after residual refinement, the final boundary prompt is

$D = \sigma(D_c + \Delta D).$

For scan direction $s$, the spatial prompt is

$$T_s = 1 + D \cdot \mathrm{ReLU}(\Phi_{dir}^{s}), \qquad F_s' = F \odot T_s.$$

The reported interpretation is boundary-sensitive propagation: high $D$ and favorable directional flow strengthen scan intensity, while propagation across irrelevant regions is diminished. On Cityscapes, DeepLabV3+ improves from 75.17 mIoU to 76.58 with G-Mamba and to 77.57 with Cascade G-Mamba; OCRNet improves from 78.22 to 79.01 and then 79.57. The added cost at $1024\times1024$ is reported as moderate, with $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$0GFLOPs ranging from +33.0 to +51.0 depending on architecture (Chan et al., 7 Jun 2026).

Remote sensing segmentation uses similar hybridization but often in encoder-decoder form. CM-UNet employs a CNN encoder for local features, a Multi-Scale Attention Aggregation module to merge different scale features, and a CSMamba decoder in which a 2D-SSM global path is gated by channel and spatial attention: $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$1 On ISPRS Potsdam, ISPRS Vaihingen, and LoveDA, the reported mIoU values are 87.21, 85.48, and 52.17, respectively (Liu et al., 2024).

Medical segmentation papers push the hybrid paradigm into 2D, 3D, and instance settings. MambaVesselNet++ uses a texture-aware convolutional layer in the Hi-Encoder and Mamba-based long-range modeling before a bifocal fusion decoder; reported results include Dice 0.870 on the IXI 3D vessel task, Dice 0.953 on PH2, and Dice 0.911 on CVC-ClinicDB (Xu et al., 26 Jul 2025). MedSegMamba places nine VSS3D blocks in the bottleneck of a 3D patch-based U-Net-like network and reports overall subcortical segmentation performance of DSC 0.88383, VS 0.97076, and ASSD 0.33604, with a significant ASSD improvement over another Mamba-based model and approximately 20% fewer parameters (Cao et al., 2024). For weakly supervised ultrasound segmentation, MambaEviScrib trains a U-Net branch and a Mamba-UNet branch jointly under Evidence-Guided Consistency, using Dempster-Shafer Theory to convert evidence $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$2 into Dirichlet parameters $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$3, belief masses, and uncertainty masses $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$4; only the U-Net branch is used at inference (Han et al., 2024).

Crack segmentation provides a more cautionary dense-prediction perspective. In one study, replacing the four middle stages’ conv blocks of CrackSeU-B with Mamba variants showed that Vim, vanilla VSS, and the VSS block from VMamba did not consistently improve performance across datasets, whereas CrackMamba did. On Deepcrack, CrackSeU-B improves from mi IoU 71.71 and mi Dice 81.40 to 73.12 and 82.50, while parameters drop from 3.19M to 1.83M and MACs from 8.94G to 7.94G (He et al., 2024). A separate VMamba-based encoder-decoder, VM-UNet, reports 27M parameters, 16G FLOPs, and 16 ms inference on $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$5 images, with mDS 85.7 on Ozgenel and performance close to or matching stronger Transformer baselines on other crack datasets (Chen et al., 2024).

Object detection extends the same local-global division. MambaNeXt-YOLO uses a ConvNeXt local branch, a Mamba global branch, ResGate fusion, and a Multi-branch Asymmetric Fusion Pyramid Network, achieving 66.6% mAP at 31.9 FPS on PASCAL VOC without any pre-training, with 7.1M parameters and 22.4G FLOPs (Lei et al., 4 Jun 2025). RemoteDet-Mamba combines Siamese CNN encoders with a Cross-modal Fusion Mamba module that flattens multimodal features along four directions, fuses them by element-wise addition, and processes them with Mamba S6 blocks; on DroneVehicle it reports 81.8 mAP, 24.01 FPS, and 71.34M size (Ren et al., 2024).

4. The facial beauty prediction model named “Mamba-CNN”

The paper explicitly titled "Mamba-CNN" defines a regression architecture for facial beauty prediction rather than a generic vision backbone (Boukhari, 1 Sep 2025). Its topology consists of a convolutional stem with $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$6 convolution, stride 2, and max pooling; a stack of MambaBlocks in strided and non-strided forms; a multi-scale feature pyramid with adaptive pooling at $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$7, $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$8, and $$B = S_B(x),\qquad C = S_C(x),\qquad A = T_A(\mathrm{Parameters} + S_A(x)),$$9; and a regression head with fully connected layers and dropout.

Its distinguishing component is not a full visual Mamba scan block, but a lightweight, Mamba-inspired gating branch embedded inside an inverted-residual-style block. The main path expands channels with $V$0 convolution, applies depthwise $V$1 convolution, and then projects back with another $V$2 convolution. In parallel, a second depthwise $V$3 convolution followed by sigmoid produces a spatial gate, yielding the content-aware recalibration

$V$4

This is described as an adaptation of content-aware selectivity from State Space Models into a convolutional context (Boukhari, 1 Sep 2025).

On SCUT-FBP5500, the model reports a Pearson Correlation of 0.9187, a Mean Absolute Error of 0.2022, and a Root Mean Square Error of 0.2610 (Boukhari, 1 Sep 2025). The same paper reports that ablations attribute a +1.0% PC improvement to the MambaBlock and a +0.7% improvement to the feature pyramid, with the full model improving PC from 0.9015 to 0.9187. This usage is important terminologically: “Mamba-CNN” here denotes a specific gating-based hybrid rather than the broader class of CNN-plus-selective-scan systems.

5. Classification, multimodal analysis, and biosignals

Outside dense prediction, Mamba-CNN hybrids appear in microscopy, medical classification, multimodal sentiment analysis, EEG decoding, ECG classification, and affective computing. In microscopic image classification, MambaMIC employs a Local-Global dual-branch aggregation module in which the input is split across channels: one half goes through depthwise separable and pointwise convolution, while the other half enters a PEVM block with ARVSS units and a Partially Selective Feed-Forward Network. Fusion is handled by the Feature Modulation Interaction Aggregation Module,

$V$5

The reported Ms-Mamba-t configuration has 4.32M parameters and reaches 88.41 RPE OA, 98.29 RPE AUC, 87.30 MHIST OA, 93.49 MHIST AUC, 97.31 SARS OA, 99.60 SARS AUC, 72.14 TissueMNIST OA, and 98.75 MedFM Colon OA (Zou et al., 2024).

In breast ultrasound image classification, a direct comparison of CNNs, ViTs, and Mamba-based encoders reports that on the B dataset the best Mamba-based models achieve a 1.98\% average AUC and a 5.0\% average Accuracy improvement compared to the best non-Mamba-based model in that study, with paired $V$6-tests at $V$7 and no cases where a CNN or ViT significantly outperforms a Mamba-based model (Nasiri-Sarvi et al., 2024). This result is narrower than a general superiority claim: it is dataset-specific and reported within a transfer-learning regime using ImageNet-pretrained models.

Multimodal sentiment analysis uses a more explicitly sequential hybridization. HCMEN first applies 1D CNNs for local unimodal modeling and then Bi-Mamba for long-range dependencies: $V$8 After Cross-Modal Enhancement and Alignment, fusion uses interleaved modality tokens and stacked hybrid blocks. On MOSI, the reported HCMEN scores are Acc-7 $V$9, F1 $\Phi$0, and MAE $\Phi$1; an ablation reports F1 values of 73.62 without CNN, 72.84 without Mamba, 72.61 without CMEA, and 74.78 for the full model (Li et al., 31 Jul 2025).

Sequence-centric biosignal tasks often use CNNs for local temporal extraction followed by Mamba for long-horizon dependency modeling. SWIM applies a short-window CNN to 1-second EEG windows and then feeds the resulting 64-dimensional features into a 3-layer Mamba backbone; in the leave-one-speaker-out setup on KUL, SW$\Phi$2 combined reaches 84.9% accuracy and SWIM reaches 86.2%, reducing classification errors by a relative 31.0% compared to the previous state-of-the-art result (Zhang et al., 2024). In 12-lead ECG, 1DCNN-ECG-Mamba uses a single 1D convolutional front-end followed by a 16-block bidirectional Vim encoder and reports AUPRC/AUROC of 0.5787/0.9571 on PhysioNet CinC 2020 and 0.6410/0.9695 on PhysioNet CinC 2021 (Jiang et al., 14 Oct 2025).

A broader implication is that Mamba-CNN is not confined to spatial image tensors. The same local-first, long-range-second decomposition appears in 1D physiological time series, multimodal token streams, and EEG window sequences, suggesting that the architectural idea is modality-agnostic even when individual implementations remain highly task-specific.

6. Performance profile, misconceptions, and open problems

A common misconception is that inserting any Mamba block into a CNN will reliably improve performance. The crack-segmentation study directly contradicts this: Vim, vanilla VSS, and the VSS block from VMamba did not consistently improve CrackSeU’s performance on both datasets, whereas the task-specific CrackMamba was the only Mamba-based block that consistently improved all evaluation metrics across both datasets (He et al., 2024). This suggests that success depends on how selective scan is adapted to the target signal structure, not merely on the presence of an SSM block.

A second misconception is that Mamba-CNN hybrids necessarily dominate strong CNN baselines. In PPG-based affect recognition under subject-independent 5-fold cross-validation, Transformer and Mamba models achieved performance comparable to a CNN baseline but did not consistently outperform it across tasks; the CNN remained the most effective overall, providing the highest accuracy with the smallest model size, while Mamba variants had 5–8× more parameters despite marginally faster inference (Alghoul et al., 28 Apr 2026). This result places a clear boundary on generalized claims of superiority, particularly for small and noisy datasets.

The survey literature identifies several unresolved design issues. Scanning strategy remains central because 2D or 3D serialization can disrupt spatial correlations, and there is no consensus on the best scan pattern across tasks. The same sources note interpretability limits, the relative scarcity of pretrained weights, and the risk of scan-dependent bias in learned representations (Rahman et al., 2024). Lightweight visual backbones such as MobileMamba address throughput by using a three-stage network, channel splitting, wavelet-enhanced Mamba, multi-kernel depthwise convolution, and identity pathways; MobileMamba reports up to 83.6% Top-1 and up to $\Phi$3 faster throughput than LocalVim on GPU, but this is a specialized solution to the practical problem that low FLOPs do not automatically translate into high throughput (He et al., 2024).

The present literature therefore supports a precise characterization. Mamba-CNN is best understood as a hybrid systems paradigm in which convolution provides locality, structural bias, and efficient early feature extraction, while Mamba or Mamba-inspired selectivity provides long-range propagation, context aggregation, or adaptive gating. Its strongest empirical results arise when the interface between those roles is designed around task structure—geometry in semantic segmentation, frequency decomposition in dehazing, bifocal fusion in medical segmentation, token-level alignment in multimodal sentiment analysis, or short-window feature extraction in EEG decoding—rather than when a generic visual Mamba block is inserted without further adaptation (Chan et al., 7 Jun 2026, Wang et al., 1 Jul 2025, Xu et al., 26 Jul 2025).