Ultralight Med-Vision Mamba Models

Updated 3 July 2026

Ultralight Med-Vision Mamba models are efficient medical vision architectures that use state-space modeling and depthwise convolutions for accurate segmentation and classification.
They integrate patch embeddings, hybrid SSM-convolution blocks, and adaptive attention modules to achieve state-of-the-art performance with significantly reduced parameters and FLOPs.
Designed for edge and point-of-care deployments, these models offer real-time inference with minimal memory footprints while addressing clinical resource constraints.

Ultralight Med-Vision Mamba refers to a class of medical vision models that leverage state-space modeling (SSM) architectures for highly efficient, low-parameter image analysis, targeting segmentation and classification tasks in clinical imaging with deployment feasibility on resource-limited devices. These architectures exploit the linear-time, long-range modeling capabilities of SSMs—particularly the Mamba variant—augmented by depthwise separable convolutions, local-global multi-branch fusion, and adaptive attention modules. Multiple variants, including HC-Mamba, UltraLight Med-Vision Mamba, UltraLBM-UNet, MambaU-Lite, and UltraLight VM-UNet, have demonstrated state-of-the-art accuracy with orders-of-magnitude fewer parameters and FLOPs than conventional CNNs or ViT backbones (Xu, 2024, Sultana et al., 12 Aug 2025, Fan et al., 25 Dec 2025, Nguyen et al., 2024, Wu et al., 2024).

1. Architectural Principles of Ultralight Med-Vision Mamba

Ultralight Med-Vision Mamba models are characterized by modular, hierarchical backbones operating on patch-based embeddings, with a strong emphasis on computational efficiency. The canonical form consists of the following:

Patch Embedding: The input image (typically $H\times W\times 3$ for 2D tasks) is split into non-overlapping patches (e.g., $4\times 4$ ), each linearly projected to a fixed number of channels, followed by LayerNorm. For 3D, analogous cube embeddings are used with dimension-independent processing (Xu, 2024, Dai et al., 19 Feb 2025).
Hybrid SSM-Convolution Blocks: The core processing involves parallel branches: one operates via a discretized Mamba SSM (often bidirectional), handling long-range dependencies; the other branch utilizes depthwise-separable and/or dilated convolutions for local context extraction (Xu, 2024, Wu et al., 2024).
Channel Split and Shuffle: Input channels are split for branchwise operations and recombined via concatenation and channel-shuffling to maximize cross-branch interaction at minimal cost.
Global-Local Multi-Branch Perception: UltraLBM-UNet, for example, decomposes features into global (bidirectional Mamba SSM), local (depthwise convolution), and identity branches, with pointwise fusion and progressive receptive field adaptation (Fan et al., 25 Dec 2025).
Attention Modulation: MambaU-Lite and others include lightweight integrated channel-spatial attention (e.g., ICSA, CBAM) and spatial/channel bridges (SCAB, CAB, SAB) to boost discriminative capacity at minimal memory overhead (Nguyen et al., 2024, Sultana et al., 12 Aug 2025, Wu et al., 2024).
Low-Rank and Quantized Operations: Efficient SSM blocks use shared, low-rank, and quantized state/transition matrices, and exploit linear scaling with input sequence length and channel dimension (Fan et al., 25 Dec 2025, Liu et al., 2024).

A representative architecture (UltraLBM-UNet) reduces to 0.034M parameters, 0.060 GFLOPs, via systematic use of these constructs, while distilled student models can reach 0.011M parameters (Fan et al., 25 Dec 2025).

2. Parameter Efficiency and Computational Complexity

The ultralight designs achieve significant reductions in both parameter count and floating-point operations relative to standard medical imaging baselines:

Model	Params (M)	GFLOPs	Notable Architecture
HC-Mamba	13.9	28	4-stage hybrid SSM-dilated-depthwise
UltraLight Med-Vision Mamba	0.049	1.2	6-stage conv+PVM, SCAB fusion
UltraLBM-UNet	0.034	0.060	Bidirect. Mamba+DwConv, GLMBP/LMBP
MambaU-Lite	0.42	1.25	U-Net w/ P-Mamba, ICSA bottleneck
UltraLight VM-UNet	0.049	0.060	4-way parallel Vision Mamba
LightM-UNet	1.09–1.87	267–457	Pure SSM UNet, RVM layers

Depthwise separable convolutions yield 8–9x parameter reduction per layer ( $K=3$ case), and combining with dilated convolutions provides ≈60% savings versus plain 3×3 convolutions (Xu, 2024).
The SSM modules guarantee $O(N)$ time and $O(1)$ per-token memory, permitting efficient inference on large images or volumetric scans (Latibari et al., 16 Feb 2026).
Four-way parallelization in UltraLight VM-UNet (PVM Layer) mathematically collapses Mamba's quadratic parameter scaling with channel dimension, achieving <0.05M parameters at full scale (Wu et al., 2024).

3. Quantitative Performance and Empirical Benchmarks

Ultralight Med-Vision Mamba models achieve or exceed the segmentation and classification accuracy of much larger CNN, Transformer, or hybrid baselines on public clinical datasets:

Skin Lesion Segmentation (ISIC17/ISIC18/PH2):
- UltraLBM-UNet: 88.78% DSC (ISIC17), 88.85% (ISIC18), 91.85% (PH2, student) (Fan et al., 25 Dec 2025).
- MambaU-Lite: 90.57% DSC (ISIC18), 95.72% (PH2) (Nguyen et al., 2024).
- UltraLight VM-UNet: 90.91% (ISIC17), 89.40% (ISIC18), 92.65% (PH2) (Wu et al., 2024).
Organ Segmentation (Synapse, AMOS, LiTs):
- HC-Mamba: 81.58% DSC (Synapse), HD95 = 26.34 mm (Xu, 2024).
- LightM-UNet: 84.58% DSC, 77.48% mIoU (LiTs); 96.17% DSC (Montgomery–Shenzhen CXR) (Liao et al., 2024).
Classification Tasks:
- UltraLight Med-Vision Mamba: 97.34% accuracy, F1 = 0.9733, 49K params on tubular adenoma progression (Sultana et al., 12 Aug 2025).
- MedMamba-Tiny: 86.2% average OA across 16 non-MedMNIST/MedMNIST datasets at 2.0 GFLOPs (Yue et al., 2024).

A direct ablation within HC-Mamba demonstrates that the union of dilated and depthwise convolutions increases mIoU by 2.75 percentage points over standard convolutions while halving parameters (Xu, 2024).

4. Design Innovations and Block-Level Mechanics

Core to the ultralight Mamba family is the exploitation of SSMs for dynamic, long-range modeling, made hardware-feasible through design choices including:

Discretized Mamba (SS2D): Implements $h_k = \bar{A} h_{k-1} + \bar{B} x_k$ , with $\bar{A},\bar{B}$ input-adaptive, and efficient convolutional scan along image axes (Yue et al., 2024, Liu et al., 2024).
Parallel Branching: Four-branch (PVM) or multi-branch (GLMBP/LMBP) modules facilitate independent SSM/dwconv processing per channel partition, with learnable re-weighted recombination (Fan et al., 25 Dec 2025, Wu et al., 2024).
Attention Integration: Prioritized channel and spatial attention (e.g., ICSA in MambaU-Lite, SCAB in UltraLBM-UNet), adaptive feature fusion with attention-based bridges (e.g., CAB/SAB), and dynamic skip-scaling (Nguyen et al., 2024, Sultana et al., 12 Aug 2025, Fan et al., 25 Dec 2025).
Dimension-Independence for 3D: Networks like MobileViM process each spatial axis separately with SSMs, maintaining linear scaling in three dimensions and enabling real-time volumetric segmentation (e.g., >90 FPS, 6.29M params) (Dai et al., 19 Feb 2025).

Block-level pseudocode and analytical formulas for parameter/FLOPs contribution are explicitly reported in key references, enabling reproducibility and informed compression (Nguyen et al., 2024, Liao et al., 2024, Yue et al., 2024).

5. Practical Applications and Deployment Feasibility

Ultralight Med-Vision Mamba models are specifically tuned for clinical environments with strict constraints:

Edge and POC Devices: Sub-1M or even sub-0.05M parameter footprints (<1 MB flash), inference latencies <10 ms on ARM/mobile SoCs, batch sizes of 1–2 for real-time operation. Quantized int8 inference yields <1% accuracy drop (Nguyen et al., 2024, Wu et al., 2024, Xu, 2024).
Workflow Integration: UltraLight Med-Vision Mamba enables real-time overlay for colonoscopy video, rapid WSIs screening, secondary radiology reads, and point-of-care dermatoscopic lesion detection (Sultana et al., 12 Aug 2025, Nguyen et al., 2024).
Clinical Accuracy vs. Resource Trade-off: On datasets such as ISIC, performance remains within ±0.2% DSC of much larger models, with >10–100× reduction in parameter scale, as plotted in explicit model-size vs. latency vs. accuracy charts (Liu et al., 2024, Wu et al., 2024).
Specialized Clinical Tasks: MRI super-resolution tasks have been addressed by efficient Mamba-based models achieving <1M params, 57 GFLOPs, and record SSIM/PSNR/LPIPS/GMSD on brain and prostate datasets, outperforming transformers and diffusion models at <1.5% of their parameter count (Safari et al., 22 Dec 2025).

6. Robustness, Limitations, and Future Directions

While ultralight Med-Vision Mamba architectures excel in clean clinical tasks, recent investigation highlights several vulnerabilities:

Adversarial Robustness: SSM models reveal pronounced susceptibility to white-box PGD/FGSM perturbations—~60–70 percentage point drops in accuracy on MedMNIST, with occlusions and noise also causing severe degradation (Latibari et al., 16 Feb 2026).
Bit-Flip and Hardware Faults: Patch embedding and SSM state matrices are sensitive to single/multiple bit flips, often amplifying errors and resulting in catastrophic failures unless error correction is deployed (Latibari et al., 16 Feb 2026).
Countermeasures: Recommendations include adversarial-aware/corruption training, spectral normalization of SSMs, local smoothing blocks, ECC-protected storage, and lightweight model ensembles (Latibari et al., 16 Feb 2026).
Extension to 3D and Multi-Modal: Active research seeks to further compress footprints (<10M or sub-1M parameters) for 3D volume tasks and to integrate explainable AI and semi-supervised learning, as well as to address small-structure segmentation failures evident in ultrasound datasets (Xu, 2024, Xu et al., 14 Jun 2025).
Resource-Accuracy Scaling: Model size and compute can be systematically traded off against target accuracy; ablations across variants quantify the minimal necessary complexity for clinical relevance (Nguyen et al., 2024, Wu et al., 2024, Dai et al., 19 Feb 2025).

7. Comparative Summary Table

Model	Task	Params (M)	FLOPs	Peak Acc.	Deployment Context
HC-Mamba (Xu, 2024)	Organ/Skin Seg.	13.9	28 G	DSC=88.18%	Edge segmenter, 224×224, 14MB model
UltraLight Med-Vision Mamba (Sultana et al., 12 Aug 2025)	CRC Progression Clf.	0.049	1.2 G	97.3% ACC	WSI, <2GB GPU, tile inference ~2ms
UltraLBM-UNet (Fan et al., 25 Dec 2025)	Skin Lesion Seg.	0.034	0.060 G	88.78% DSC	0.011M student variant, point-of-care
MambaU-Lite (Nguyen et al., 2024)	Skin Lesion Seg.	0.42	1.25 G	95.72% DSC	≈1.67MB, real-time on CPU/mobile
UltraLight VM-UNet (Wu et al., 2024)	Skin Lesion Seg.	0.049	0.060 G	90.91% DSC	200KB FP16, mHealth/edge FPGA
LightM-UNet (Liao et al., 2024)	Organ/Tumor Seg.	1.1–1.9	267–457 M	96.17% DSC	16MB memory, int8 quantized
MobileViM (Dai et al., 19 Feb 2025)	3D Med. Seg.	6.29	196 G	Dice=92.7%	91 FPS RTX4090, <25MB

Peak accuracy and resource requirements are dataset and input-size dependent.

These models realize a new efficiency-accuracy Pareto frontier for medical vision, grounded in SSM theory, parallel channel-batch computation, and modular hybridization of local and global feature processing. Their empirical success and practical deployability position them as leading tools for resource-constrained clinical machine vision.