MambaVesselNet++: Hybrid CNN–Mamba Segmentation

• The paper introduces a hybrid CNN–Mamba architecture that combines local texture modeling with efficient long-range dependency capture for medical image segmentation.
• It employs a Hi-Encoder for multi-scale feature extraction and a BF-Decoder that fuses global context using selective state-space modeling to enhance precision.
• Experimental results on six public datasets demonstrate superior Dice scores, reduced memory footprint, and faster training compared to transformer and CNN baselines.

Searching arXiv for the specified paper and closely related context papers to ground the article. MambaVesselNet++ is a hybrid CNN–Mamba architecture for medical image segmentation that follows a U-shaped encoder–decoder design and is intended to combine local texture modeling with efficient long-range dependency capture (Xu et al., 26 Jul 2025). The model is composed of a hybrid image encoder (Hi-Encoder) and a bifocal fusion decoder (BF-Decoder). Within this design, convolutions are used to extract multi-scale, low-level semantic features, while selective state-space modeling via Mamba is used to model global context with linear complexity. The reported scope includes 2D semantic segmentation, 2D instance segmentation, and 3D semantic segmentation, with experiments on PH2, CVC-ClinicDB, DRIVE, UDIAT, TNBC, and IXI (Xu et al., 26 Jul 2025).

1. Architectural definition and design rationale

At a high level, MambaVesselNet++ follows the familiar encoder–decoder paradigm. The Hi-Encoder contains two stages: texture-aware layers for extracting low-level semantic features and progressively downsampling the input, followed by Vision Mamba layers for modeling global, long-range dependencies with linear complexity. The BF-Decoder then performs hierarchical upsampling with deconvolutions and skip-connections that fuse local texture features from the encoder with global context features from the state-space backbone in a bifocal manner (Xu et al., 26 Jul 2025).

The architecture is defined for both 2D images and 3D patches. The overview given in the paper is:

$2\times 2$5

The design places pure convolutional layers at the edges, where precise spatial detail matters most, and Mamba blocks at the bottleneck, where global context matters most. This arrangement is presented as a balance between two limitations identified in prior families of models: convolution-based U-shape architectures are usually limited by the local receptive field, while vision transformers, despite strong global modeling, are challenged in real-world deployment by the non-linear self-attention mechanism and its computational cost (Xu et al., 26 Jul 2025). A plausible implication is that the model is explicitly structured to avoid using a globally contextual mechanism at resolutions where boundary fidelity is most vulnerable.

2. Hi-Encoder and the texture-aware layer

The first half of each encoder stage is a texture-aware block designed to extract local patterns, including edges, corners, and textures, while downsampling. If the input feature map at layer $i$ is denoted by

$$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$

the layer applies:

1. a $3\times 3$ convolution with weight $W^{(3)}_{i}$ and bias $b^{(3)}_{i}$,
2. LayerNorm,
3. LeakyReLU,
4. a $2\times 2$ convolution with stride $2$, and
5. residual addition.

The compact formulation is

$$\tilde f_i = \mathrm{LeakyReLU}\Bigl(\mathrm{LN}\bigl(W^{(3)}_{i} * f_{i-1} + b^{(3)}_{i}\bigr)\Bigr), \quad,\quad f_i = \mathrm{LeakyReLU}\Bigl(\mathrm{LN}\bigl(W^{(2)}_{i} * \tilde f_i + b^{(2)}_{i}\bigr)\Bigr) + \tilde f_i,$$

where “$*$” denotes convolution and the second $2\times 2$ convolution reduces spatial size by half while doubling the channel count (Xu et al., 26 Jul 2025).

A specific claim in the paper is that, by contrast to max-pooling, the strided convolution preserves more low-frequency information. In architectural terms, the texture-aware layer therefore performs both feature extraction and resolution reduction without introducing a separate pooling operator. This suggests that the encoder is intended to preserve local signal structure before the tokenization and sequence modeling steps.

After $$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$0 texture-aware downsampling stages, the feature tensor

$$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$1

is reshaped into a sequence of

$$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$2

tokens

$$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$3

which then enter the Vision Mamba portion of the encoder (Xu et al., 26 Jul 2025).

3. Selective state-space modeling and Vision Mamba integration

The Mamba component is introduced through the state-space model formalism. A continuous-time SSM is written as

$$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$4

where $$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$5 is the hidden state, $$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$6 is the input, and $$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$7 are learned matrices. After discretization, this yields a linear recurrence computable in $$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$8 time (Xu et al., 26 Jul 2025).

Mamba extends this formulation by dynamic gating, termed selective scanning, over the input and state updates. For block $$f_{i-1}\in\mathbb{R}^{C_{i-1}\times H_{i-1}\times W_{i-1}\,(\times D_{i-1})},$$9,

$3\times 3$0

Each Mamba call is defined as a linear scan over the sequence: $3\times 3$1 where $3\times 3$2 is a selection gate, $3\times 3$3 are gating parameters, and the convolutional-S6 kernels $3\times 3$4 are shared across the scan (Xu et al., 26 Jul 2025).

The paper explicitly contrasts this with self-attention. For $3\times 3$5 tokens of dimension $3\times 3$6,

$3\times 3$7

with cost $3\times 3$8, whereas the Mamba recurrence costs $3\times 3$9 (Xu et al., 26 Jul 2025). This complexity claim is central to the model’s positioning relative to transformer-based segmentation backbones. The paper does not state that Mamba fully supersedes attention in all settings; rather, it argues that selective SSMs provide a more tractable mechanism for long-range dependency modeling in medical segmentation.

4. BF-Decoder and bifocal fusion

The bifocal fusion decoder reconstructs segmentation masks by progressively upsampling the global feature representation and reintroducing encoder features through skip connections. If $W^{(3)}_{i}$0 Mamba blocks produce a global feature tensor

$W^{(3)}_{i}$1

the deepest decoder stage is

$W^{(3)}_{i}$2

where $W^{(3)}_{i}$3 is element-wise addition and the deconvolution upsamples spatially by a factor of $W^{(3)}_{i}$4 (Xu et al., 26 Jul 2025).

For shallower stages $W^{(3)}_{i}$5,

$W^{(3)}_{i}$6

where $W^{(3)}_{i}$7 denotes channel-wise concatenation with the corresponding encoder feature. A final $W^{(3)}_{i}$8 convolution maps $W^{(3)}_{i}$9 to the desired number of output channels, followed by Sigmoid or Softmax (Xu et al., 26 Jul 2025).

The paper also gives the following pseudocode:

$2\times 2$6

The term “bifocal” refers to the combination of local texture features and global context features. In operational terms, the decoder does not simply mirror the encoder; it combines additive fusion at the deepest stage with concatenative fusion at subsequent scales. This suggests an attempt to preserve bottleneck-level global semantics while reintroducing progressively finer local cues.

5. Training protocol and task coverage

The reported datasets and tasks are divided into 2D semantic, 2D instance, and 3D semantic segmentation. The 2D semantic datasets are PH2 for dermoscopy, CVC-ClinicDB for polyp segmentation, DRIVE for fundus imaging, and UDIAT for ultrasound. The 2D instance dataset is TNBC for histopathology nuclei. The 3D semantic dataset is IXI MRA for cerebrovascular segmentation (Xu et al., 26 Jul 2025).

Data splits are given as $b^{(3)}_{i}$0 for 2D datasets and official splits for DRIVE and IXI. Optimization uses Adam with initial learning rate $b^{(3)}_{i}$1 and Cosine Annealing down to $b^{(3)}_{i}$2. For 3D training, the setup is 5,000 iterations, batch size 2 per GPU, and patch size $b^{(3)}_{i}$3. For 2D training, the setup is 200 epochs with batch size 16. Augmentations are standard random flips, intensity jitter, and rotations as in MONAI defaults (Xu et al., 26 Jul 2025).

For semantic segmentation, the loss is

$b^{(3)}_{i}$4

with

$b^{(3)}_{i}$5

and $b^{(3)}_{i}$6 (Xu et al., 26 Jul 2025).

For the 3-head instance setting, the loss is

$b^{(3)}_{i}$7

where

$b^{(3)}_{i}$8

with $b^{(3)}_{i}$9 (Xu et al., 26 Jul 2025).

6. Experimental performance, ablations, and computational profile

The paper reports that, across six public datasets, MambaVesselNet++ outperforms current convolution-based, transformer-based, and Mamba-based state-of-the-arts (Xu et al., 26 Jul 2025). The comparisons include UNet3D and nnUNet as purely convolutional baselines, TransUNet, UNETR, and SwinUNETR as transformer-based baselines, and U-Mamba, SegMamba, and Swin-U-Mamba as SSM-based baselines.

The headline quantitative results are as follows:

Task / dataset	Metrics
PH2	Dice 0.953 / mIoU 0.911 / HD 19.50
CVC-ClinicDB	Dice 0.911 / mIoU 0.854 / HD 13.22
DRIVE	Dice 0.711 / mIoU 0.552 / HD 18.12
UDIAT	Dice 0.849 / mIoU 0.763 / HD 23.01
TNBC	AJI 0.534 / PQ 0.479 / DQ 0.657 / SQ 0.729
IXI	Precision 0.889 / Recall 0.859 / Dice 0.870

The paper states that MambaVesselNet++ consistently yields the highest Dice/mIoU and lowest Hausdorff. In 3D, it also reduces per-epoch training time and memory relative to SwinUNETR, specifically 2.1 min versus 3.8 min and 6.9 GB versus 9.8 GB (Xu et al., 26 Jul 2025). Domain generalization tests, including PH2 $2\times 2$0 ISIC2018 and DRIVE $2\times 2$1 STARE, are reported to further confirm robustness under unseen imaging conditions.

The ablation study on the IXI dataset isolates the contribution of each major component. Starting from a UNETR baseline with Dice 0.836, adding the texture-aware layer yields 0.840, adding the Vision Mamba layer yields 0.851, the full Hi-Encoder yields 0.859, and the combination of Hi-Encoder and BF-Decoder yields 0.870 (Xu et al., 26 Jul 2025). Within the limits of the reported study, this progression attributes the largest total gain to the combined hybrid encoder and bifocal decoder rather than to any single component in isolation.

The complexity analysis compares convolution, self-attention, and Mamba in Big-O form: $2\times 2$2

$2\times 2$3

$2\times 2$4

The empirical interpretation given is that restricting Mamba blocks to the bottleneck, rather than placing them at every stage, keeps the overall cost dominated by cheap strided convolutions and small MLPs, producing a net speed and memory advantage over full Vision Transformer backbones (Xu et al., 26 Jul 2025).

7. Limitations, interpretive cautions, and future directions

The paper explicitly notes several limitations and trade-offs. One is the risk of over-segmentation when global context overwhelms local edges; this is mitigated by limiting Mamba to the bottleneck. Another is under-segmentation in purely CNN methods, which the architecture seeks to avoid through SSM integration (Xu et al., 26 Jul 2025). These points serve as a caution against a simplistic reading that more global context is always preferable. In the reported framing, the central issue is not merely adding global modeling capacity, but controlling where and how it is introduced.

The listed future directions are adaptive placement of Mamba blocks at multiple scales, joint learning of selection gates for better interpretability, and extension of the multi-branch decoder to panoptic tasks (Xu et al., 26 Jul 2025). These proposals indicate that the current model uses a deliberately constrained placement strategy for state-space layers and that interpretability of the selection mechanism remains an open concern.

Taken as a whole, MambaVesselNet++ is presented as a judicious hybrid of CNN feature extraction and selective SSM sequence modeling. The main technical claim is not that either component is sufficient on its own, but that the architectural separation of local-detail extraction, bottleneck-level long-range modeling, and bifocal decoder fusion yields strong segmentation performance while keeping computational and memory costs tractable (Xu et al., 26 Jul 2025).