Selective State-Space Models in Medical Image Processing

1. Introduction

Modeling local and global features in high resolution medical images presents a significant challenge. Although Convolutional Neural Networks (CNNs), one of the existing methods, successfully extract local details, they struggle to capture long-range dependencies between distant pixels (Jiang et al., 2026; C. Li et al., 2025). On the other hand, since the computational cost of Transformer (ViT) architectures that can capture global dependencies increases quadratically with image size, their use in high-resolution medical imaging is severely limited (Jiang et al., 2026; Y. Li et al., 2026).

In existing studies, approaches such as attention mechanisms or dividing images into small patches are used to overcome these limitations (Jiang et al., 2026; Zhu et al., 2024). Recently, Selective State Space Models (Selective SSM - Mamba) with linear computational cost are developed as an alternative to overcome complexity and prevent local-global information loss (Jiang et al., 2026; C. Li et al., 2025). The Mamba architecture was designed for processing images without bottlenecking and at high speed. In contrast to other deep learning architectures, it has a data-dependent selective state memory (Jiang et al., 2026; Zhu et al., 2024). This mechanism filters out unnecessary spaces and noises in the images ensuring that only critical anatomical or pathological areas that need to be focused on are kept in memory (Gu & Dao, 2024; C. Li et al., 2025). Recent studies in literature prove that architectural approaches in medical field show a strong trend from CNNs to ViTs and finally to Mamba (Ruan et al., 2024). In this context, hybrid models are developed by integrating convolutional layers, attention modules, and ViTs between Mamba blocks to prevent information loss (C. Li et al., 2025; Y. Li et al., 2026; Zhou et al., 2026).

Furthermore, there is a growing focus on lightweight designs with reduced number of parameters that can operate in mobile platforms or limited resources and various scanning mechanisms that assure preservation of long-range dependencies between distant pixels (Jiang et al., 2026; Y. Li et al., 2026; Zhou et al., 2026).

2. Methodology

2.1. Mamba Architecture

State Space Model (SSM), is a mathematical framework developed to efficiently process sequence data, inspired by continuous dynamic systems (Wang et al., 2025; Yue & Li, 2024). Given the limitations of traditional models and Transformers architectures, development of Mamba, which is a selective SSM, brings out a strong alternative to these limitations. The fundamental operating principle of Mamba relies on a data-dependent selective mechanism that dynamically updates parameters based on input sequences (Gu & Dao, 2024; Ruan et al., 2024; Yue & Li, 2024; Zhu et al., 2024). This mechanism allows the model to filter out noises and retain only crucial information in its memory to ensure high performance on long sequences. In addition, Mamba architecture optimizes GPU memory hierarchy (SRAM and HBM) by using a hardware-aware algorithm. Thus, linear computational complexity is maintained while training and inference processes are speeding up (Gu & Dao, 2024; Hedhoud et al., 2025; Ruan et al., 2024; Yurdusever et al., 2025). Scheme of Mamba (Vision Mamba – bidirectional scan) architecture is given in Figure 1 (Zhu et al., 2024).

When it comes to the mathematical basis, Mamba uses continuous-time ordinary differential equations that map one-dimensional input sequence $\left(x\right(t\left)\right)$ to output sequence $\left(y\right(t\left)\right)$ through a hidden state $\left(h\right(t\left)\right)$. This relation is demonstrated via Equation (1) (Guo et al., 2025; Yue & Li, 2024).

$$\begin{array}{c}{h}^{\text{'}}\left(t\right)=Ah\left(t\right)+Bx\left(t\right)\end{array}$$

(1)

$$\begin{array}{c}y\left(t\right)=Ch\left(t\right)\end{array}$$

(2)

In Equations (1) and (2), $A$ denotes the state matrix while $B$ and $C$ denote projection parameters. This continuous system needs to be discretized to integrate it to Deep Learning environment. The Zero-Order Hold (ZOH) method is usually used in this regard (Ruan et al., 2024; Yue & Li, 2024; Yurdusever et al., 2025; Zhu et al., 2024).

$$\begin{array}{c}\overline{A}=\mathit{exp}\left(\Delta A\right)\end{array}$$

(3)

$$\begin{array}{c}\overline{B}={\left(\Delta A\right)}^{-1}(\mathit{exp}\left(\Delta A\right)-I)\cdot \left(\Delta B\right)\end{array}$$

(4)

A timeline parameter $(\Delta )$ is added to the system via Equations 3 and 4. These are used to transform continuous $A$ and $B$ matrices into discrete forms. Finally, the system calculates the next state and output via discrete iteration formulas. The main difference of Mamba from other models is that the variables $\Delta$, $B$ and $C$ are designed as functions dependent on time and input (Guo et al., 2025; Yue & Li, 2024; Zhu et al., 2024).

2.2. Literature

The literature presented within the scope of this study is created by searching fundamental scientific databases such as Scopus, IEEE, Web of Science and ACM. During the search, the focus was the state-of-the-art studies of Mamba architectures in medical imaging published between 2024 and 2026 (Jiang et al., 2026; Y. Li et al., 2026; Ruan et al., 2024).

3. Findings

Integration of Selective State Space Models (Mamba) into medical imaging area allows researchers to overcome the computational bottleneck in processing high-resolution images and resulted in the establishment of a new generation of clinical applications. The studies inspected illustrated that this innovative architecture is not limited to a specific organ or data type. Mamba architecture is successfully applied in a wide range of fields such as neurological analyses, dermatological examinations, cellular-level pathology diagnoses, and radiological scans. In this section, current Mamba-based studies in literature are analyzed. The studies are divided into two main groups: clinical application areas and architectural approaches.

3.1. Clinical Application Areas

Mamba architecture provides specific and versatile solutions for many sub-branches of medical imaging. Tumor classification using MRI data in neurology and lesion detection using dermatoscopic images are examples of primary research domains. In the lungs, tuberculosis diagnosis and lesion analysis are performed while cell and parasite classification are performed in microscopic examinations. In addition, studies such as multi-organ segmentation and modality synthesis that allow transition between imaging types are noteworthy in literature. The clinical focus, data types used and key achievements of the studies examined are summarized in Table 1.

3.2. Architectural Approaches

There are some novel approaches in literature to prevent data loss while adapting Mamba’s 1D structure to 2D and 3D medical images. These approaches have been evaluated under three main headings: hybrid architectures, advanced scanning mechanisms, and cross-paradigm integrations. Hybrid architectures provide high accuracy with lower parameters by combining CNN’s local feature extraction capability with Mamba’s global context modeling. Advanced scanning mechanisms use multi-path or dynamic scanning algorithms to analyze undirected structure of medical images while preserving spatial integrity. Finally, in cross-paradigm integration studies, Mamba architecture is combined with other powerful models such as Generative Adversarial Networks (GAN), Diffusion and Kolmogorov-Arnold (KAN). The main purpose of these studies is to improve performance and reduce memory requirements in challenging situations. Methodological innovations of the studies inspected are given in Table 2.

4. Discussion and Conclusions

Although Mamba architecture offers an important advantage by reducing the quadratic computational cost of ViTs to linear level, there are some critical problems in clinical uses. The most fundamental architectural bottleneck is the local information loss during transformation of 2D and 3D medical data to 1D sequences. While researchers attempt to address this issue by using hybrid and multi-path scanning approaches, this can sometimes overshadow Mamba’s main promise of reduced number of parameters.

On the other hand, the model’s robustness and generalization capabilities are not yet fully optimized for real-world clinical applications. The findings of literature demonstrate that Mamba models may be vulnerable to sensor noises, targeted manipulations, and hardware errors. Furthermore, domain shifts in data distribution between “data obtained from different devices in different hospitals” a common occurrence in clinical practice. It negatively affects the cross-dataset performance of the model.

Future research is expected to progress towards lightweight Mamba models focused on edge computing that are directly optimized for the limited hardware capacities of clinical devices. At the same time, it is anticipated that multi-modality synthesis and few-shot learning strategies will be used in a much more integrated way with Mamba infrastructure to overcome lack of data problems for rare diseases.

In this review, recent methodological studies and clinical applications conducted using Selective State Space Models (Mamba) in medical imaging field are examined. Comprehensive analyses illustrate that Mamba architecture offers high accuracy and memory efficiency in a wide range of fields from neurology to dermatology. It is also seen that structural information loss is significantly reduced due to hybrid approaches and dynamic scanning mechanisms. Despite security and generalization challenges, Mamba architecture has the potential to serve as a cornerstone for future generations of clinical decision support systems thanks to its linear computational advantage and capability of modeling global context.