MambaEye Visual Models

• MambaEye is a state-space modeled visual system that processes arbitrary-sized inputs using causal or bidirectional sequential methods for efficient, real-time analysis.
• It employs novel relative move embeddings and a diffusion-inspired loss to progressively build prediction certainty, outperforming traditional transformer-based models in scalability.
• The design extends to medical imaging with specialized models like MSV-Mamba and OCTAMamba, achieving state-of-the-art segmentation accuracy with linear computation and memory costs.

MambaEye is a class of visual models based on state-space modeling paradigms, distinguished by input-size agnostic architectures, causal or bidirectional sequential processing, and linear time/memory complexity relative to the number of input tokens. The term “MambaEye” encompasses both the canonical “MambaEye” encoder for arbitrary-sized vision input (Choi et al., 25 Nov 2025) and a family of Mamba-based medical imaging models in segmentation contexts (notably echocardiography (Yang et al., 13 Jan 2025) and retinal OCTA (Zou et al., 2024)). MambaEye models are characterized by their ability to efficiently process high-resolution visual data with limited computational resources while providing strong inductive biases for translation invariance, making them adaptable across modalities and scales.

1. State-Space Modeling Framework and Mamba2 Backbone

At the core of MambaEye models is the state-space model (SSM) backbone, formalized as follows: $$\frac{d\,h(t)}{dt} = A\,h(t) + B\,x(t), \qquad y(t) = C\,h(t)$$ where $h(t)$ is the latent state, $x(t)$ the input, and $y(t)$ the output. Discretization with step size $\Delta$ yields a recurrence: $$h_t = \bar{A} h_{t-1} + \bar{B} x_t,\qquad y_t = C h_t$$ The Mamba2 “pure SSM” variant adaptively selects among multiple state coefficients for each block, enhancing information flow and expressivity. This recurrence is strictly causal in the original MambaEye design, enabling the model to emit predictions at any point in a sequential scan of visual input (Choi et al., 25 Nov 2025).

Medical models such as MSV-Mamba and OCTAMamba use similar SSM blocks (possibly bidirectional (Yang et al., 13 Jan 2025); (Zou et al., 2024)) within U-shaped encoder–decoder networks to efficiently encode and fuse multi-scale spatial dependencies.

2. Causal Visual Encoding and Relative Move Embeddings

MambaEye introduces a novel size-agnostic, causal encoder for image data by flattening an image into a sequence of patches. At each timestep $t$, the encoder receives:

• The patch vector $\mathbf{v}_t$,
• A relative move embedding $\mathbf{m}_t$, encoding the shift from the previous patch via sinusoidal position encoding of $(\Delta x_t, \Delta y_t)$,
• An information ratio $r_t$, quantifying cumulative input coverage.

These quantities are concatenated, linearly projected, and processed through $L$ causal SSM blocks. The relative move embedding makes the model fully translation-invariant and agnostic to image resolution and scanning order. Arbitrary scan paths—random, learned, or fractal—are supported by design, enabling robust adaptation to diverse vision tasks without architectural modification (Choi et al., 25 Nov 2025).

3. Loss Formulation: Diffusion-Inspired Supervision

To train the model to make increasingly confident predictions as more visual evidence accrues, MambaEye incorporates a diffusion-inspired loss function. At timestep $t$, the supervised target label is interpolated between a uniform prior ($\mathbf{p}_{\rm prior}$) and the sharp one-hot class label ($\mathbf{p}_{\rm target}$), weighted by $r_t$: $$\mathbf{p}_{\rm sched}^{(t)} = (1 - r_t) \mathbf{p}_{\rm prior} + r_t \mathbf{p}_{\rm target}$$ Dense stepwise cross-entropy loss is applied over all timesteps, encouraging the model to build prediction certainty sequentially. This loss significantly improves accuracy compared to standard “sharpened” stepwise cross-entropy, particularly at high resolutions (Choi et al., 25 Nov 2025).

4. Linear Complexity Scalability

A principal advantage of MambaEye is linear scaling:

• Per-patch computation: $O(L d_{\rm model}^2)$,
• Total complexity: $O(N)$, with $N$ the number of patches.
• Inference memory: $O(1)$, as only the current SSM hidden states need be retained.

This is in contrast to transformer-based self-attention architectures, for which computation and memory scale as $O(N^2)$ due to full-pairwise attention. The linear scaling of MambaEye and its medical imaging derivatives enables efficient real-time processing of high-resolution data and is a critical factor in both vision and clinical deployment (Choi et al., 25 Nov 2025); (Yang et al., 13 Jan 2025); (Zou et al., 2024).

5. Domain-Specific Medical Models: Segmentation Architectures

The “MambaEye” design has been extended to specialized medical applications, resulting in models such as MSV-Mamba for echocardiography (Yang et al., 13 Jan 2025) and OCTAMamba for retinal vasculature segmentation (Zou et al., 2024). These systems share core SSM blocks and apply additional domain-attuned architectural strategies:

• In MSV-Mamba, cascaded residual encoders facilitate deep feature extraction, and large-window multiscale Mamba (LMS) modules in the decoder capture global context with linear complexity. Hierarchical auxiliary losses and dual attention fusion (spatial and channel branches) further improve morphologically precise segmentation performance.
• OCTAMamba integrates a Quad Stream Efficient Mining Embedding for local multi-branch features, a Multi-Scale Dilated Asymmetric Convolution Module to extract capillary-to-macro vessel structure, and Focused Feature Recalibration for robust denoising and signal enhancement. Attention gates suppress skip-connection artifacts and reinforce anatomical precision.

Both networks achieve state-of-the-art Dice accuracy and IoU in their respective benchmarks (EchoNet-Dynamic, CAMUS, OCTA-3M/6M, ROSSA), with parameter counts substantially below competing transformer or CNN baselines.

6. Empirical Results and Ablation Evidence

On ImageNet-1K, MambaEye achieves strong top-1 classification accuracy across input resolutions:

• MambaEye-B (21M params): 72.2% at $224^2$, 75.0% at $512^2$, 73.7% at $1024^2$, 66.7% at $1536^2$ (Choi et al., 25 Nov 2025).

For echocardiographic segmentation:

• MSV-Mamba: Dice of 92.92% (EchoNet-Dynamic), LV endocardium Dice of 95.01 (ED) / 93.36 (ES), LV epicardium 87.35 (ED) / 87.80 (ES), all exceeding prior Mamba or CNN-based models (Yang et al., 13 Jan 2025).

For OCTA vasculature:

• OCTAMamba: Dice/F1 on OCTA_3M is 84.5%, compared to 79.4% (U-Net) or 80.4% (AC-Mamba), with only 3.6M parameters (Zou et al., 2024).

Ablation studies demonstrate that both the causal SSM sequence modeling and translation-invariant moves (general MambaEye), as well as multiscale Mamba modules and dual-attention fusion (MSV-Mamba, OCTAMamba) are essential to maximal performance in their application domains.

7. Future Directions and Extensions

One future direction outlined in the MSV-Mamba paper is the extension to volumetric (3D) models, noted as “MambaEye‐3D,” for temporal or spatial beat-to-beat segmentation refinement and real-time dynamic modeling in cardiology (Yang et al., 13 Jan 2025). Clinical integration is anticipated through the addition of downstream diagnostics (ejection fraction, wall thickness) into MambaEye-based pipelines.

A plausible implication is that the core Mamba state-space approach, with its linear scaling and causality, will continue to see broader adoption in vision systems where efficiency, robustness to scaling, and domain-adaptive inductive bias are required.

---

References:

• MambaEye: (Choi et al., 25 Nov 2025)
• MSV-Mamba: (Yang et al., 13 Jan 2025)
• OCTAMamba: (Zou et al., 2024)