💻 computer science

Selective State Space Models in Medical Imaging: A Systematic Review of Mamba-Based Architectures and Clinical Applications

This systematic review of 445 recent studies (2024–2026) identifies CNN-Mamba hybrids as the most efficient architecture for medical imaging due to their linear computational complexity, while highlighting critical challenges like spatial information loss and memory consumption that must be addressed through native 3D blocks to enable clinical integration.

Original authors: Ali Emre Gök, Mustafa Yurdakul, Şakir Taşdemir

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali Emre Gök, Mustafa Yurdakul, Şakir Taşdemir

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Too Big to Fit" Problem

Imagine you are trying to read a massive, 3D encyclopedia of the human body (like a full MRI scan) to find a tiny tumor.

For a long time, doctors used CNNs (Convolutional Neural Networks). Think of these as a flashlight. They are great at scanning small, local areas to see fine details (like the texture of a cell), but they can't see the whole room at once. They miss the big picture.

Then came Transformers (ViT). These are like super-eyes that can see the entire room at once, understanding how the ceiling connects to the floor. However, there's a catch: the more details you try to see, the more energy they consume. If you try to look at a high-resolution 3D scan with a Transformer, it's like trying to run a marathon while carrying a heavy backpack. The computer runs out of memory (the "backpack" gets too heavy), and the process grinds to a halt.

The New Hero: Mamba

Enter Mamba (and its family, Selective State Space Models). The paper describes Mamba as a smart, efficient librarian.

Instead of trying to memorize every single book in the library at once (which takes too much space), the librarian walks down the aisles one by one. But here's the magic: the librarian is selective. As they walk, they decide instantly what information to keep in their short-term memory and what to forget.

The Benefit: They can process a massive library (a huge 3D medical scan) without needing a giant warehouse (expensive computer hardware). Their speed stays linear (steady) no matter how big the library gets, whereas the "super-eyes" (Transformers) get exponentially slower and heavier.

What the Paper Actually Did

The authors didn't invent a new Mamba; they acted as detectives. They gathered 445 recent studies (published between 2024 and 2026) to see how Mamba is being used in medicine. They sorted these studies into four main "teams" or strategies:

The Hybrid Team (CNN-Mamba): This is the most popular team. They combine the Flashlight (CNN) for local details with the Librarian (Mamba) for the big picture.
- Analogy: It's like having a detective who uses a magnifying glass for fingerprints but also has a map of the whole city.
- Result: This approach is currently the "champion" in the literature, working well for brain tumors, skin lesions, and eye scans.
The Pure Mamba Team: These researchers try to use only the Librarian, without the flashlight.
- Analogy: They are trying to read the whole book without ever stopping to look at a single word closely.
- Result: It's very fast and efficient for huge 3D data, but sometimes it misses tiny, fine details because it flattens the 3D image into a 1D line.
The Transformer-Mamba Team: These are trying to mix the Super-Eyes and the Librarian.
- Analogy: They are trying to build a robot that has both super-vision and a selective memory.
- Result: It works very well for complex tasks, but the robot is often too heavy and expensive to run on standard hospital computers.
The New Scanners Team: Since Mamba was built for 1D lists (like text), using it on 2D or 3D images is tricky. These researchers invented new ways to "scan" the image.
- Analogy: Instead of just reading a book from left to right, they invented ways to read it diagonally, in spirals, or from all four corners at once to make sure no page is missed.

The Good News

The paper claims that Mamba is a game-changer because it solves the "backpack problem." It allows computers to process high-resolution, 3D medical images (like full brain scans) much faster and with less memory than the old Transformer models. It is becoming the new standard for handling big medical data.

The Bad News (The Bottlenecks)

Despite the excitement, the paper points out three major hurdles stopping Mamba from being used in every hospital tomorrow:

The "Flattening" Problem: To make Mamba work, we often have to squash a 3D cube of an organ into a flat 1D line (like unrolling a carpet).
- The Issue: When you unroll a carpet, the neighbors that were right next to each other in the 3D world might end up far apart in the line. This causes the model to lose some "spatial sense" of how tissues connect.
The "Heavy Backpack" Problem (Memory): Even though Mamba is efficient, the popular "Hybrid" models (mixing CNNs and Mamba) still require a lot of memory.
- The Issue: Many hospitals have older computers. These models are often too heavy to run in real-time during surgery or on portable ultrasound machines.
The "One-Size-Fits-All" Problem: Most studies tested these models on public, perfect datasets (like a clean, standard library).
- The Issue: Real hospitals have messy data (noisy ultrasound, different machine types, missing scans). The paper notes that these models haven't been tested enough on real-world, messy data from different hospitals to prove they are reliable enough for critical decisions.

The Conclusion

The paper concludes that Mamba is not just a passing trend; it is likely the backbone of the next generation of medical AI. However, before it can be fully trusted in the clinic, researchers need to:

Build "Native 3D" versions that don't need to flatten images into lines.
Make the models lighter so they fit on smaller devices.
Test them on messy, real-world data from many different hospitals.

In short: Mamba is the smart, efficient librarian that can finally read the whole medical encyclopedia without getting tired, but we still need to teach it how to navigate the messy, real-world library without losing its way.

Technical Summary: Selective State Space Models in Medical Imaging

1. Problem Statement

Medical image analysis faces a critical trade-off between modeling local features and capturing global context. While Convolutional Neural Networks (CNNs) excel at local feature extraction due to their inductive bias and spatial locality, they struggle with long-range dependencies inherent in large anatomical structures. Conversely, Vision Transformers (ViTs) address global context through self-attention mechanisms but suffer from quadratic computational complexity ( $O(N^2)$ ) relative to the number of tokens. This quadratic scaling creates severe hardware and memory bottlenecks when processing high-resolution 2D images and, more critically, 3D volumetric medical data (e.g., MRI, CT), hindering their clinical deployment.

Although Selective State Space Models (Selective SSMs), specifically the Mamba architecture, offer a solution by processing sequential data with linear computational complexity ( $O(N)$ ), their direct application to medical imaging presents new challenges. Standard Mamba architectures are designed for 1D sequences; flattening 2D or 3D medical images into 1D sequences disrupts spatial continuity, topological consistency, and directional dependencies. Furthermore, adapting these models to 2D/3D data often requires multi-directional scanning mechanisms that increase memory consumption, potentially negating the efficiency gains over Transformers.

2. Methodology

This study is a systematic review conducted in accordance with PRISMA guidelines to evaluate the integration of Mamba-based architectures in medical artificial intelligence.

Search Strategy: A comprehensive search was performed across four databases (Web of Science, IEEE Xplore, Scopus, ACM Digital Library) for peer-reviewed journal articles published between 2024 and 2026.
Selection Criteria: The review focused on studies utilizing Mamba or Selective SSMs for medical/biomedical imaging tasks (segmentation, classification, synthesis, etc.). Non-peer-reviewed works, non-medical applications, and studies lacking substantial methodological contributions were excluded.
Data Synthesis: From an initial pool of 1,259 articles, 445 studies met the inclusion criteria. A representative subset of 60 studies (15 per architectural category) was selected for detailed qualitative synthesis.
Qualitative Critical Appraisal: Instead of a quantitative meta-analysis (which was deemed infeasible due to the heterogeneity of tasks and metrics), the authors employed a qualitative appraisal focusing on three criteria:
1. Methodological transparency and reproducibility (open source code, comparison to SOTA).
2. Dataset robustness and validation (public vs. private, cross-validation).
3. Computational feasibility and clinical applicability (parameter load, memory bottlenecks, generalization).

The 445 included studies were categorized into four architectural typologies:

Pure Mamba Adaptations: Direct application of Mamba backbones without hybridization.
CNN-Mamba Hybrids: Combining CNNs for local features with Mamba for global context.
Transformer-Mamba Integrations: Hybridizing ViT/Swin-Transformer blocks with Mamba.
Novel Scanning Mechanisms: Innovations in scanning paths (e.g., cross-scan, fractal, bidirectional) to preserve spatial topology.

3. Key Contributions and Findings

3.1 Architectural Dominance

The review identifies CNN-Mamba Hybrids as the most dominant and efficient approach in the current literature. These models successfully combine the local feature extraction capabilities of CNNs with the global context modeling of Mamba. They are particularly prevalent in tasks requiring both micro-level tissue detail and macro-level organ topology, such as brain tumor segmentation, skin lesion detection, and pathology analysis.

3.2 Modalities and Tasks

Modalities: The architecture shows the highest adaptability to MRI and CT (18% and 15% of studies, respectively) and multi-modality data (27%). The linear complexity of Mamba provides a natural advantage for processing the high-resolution, 3D volumetric nature of these modalities.
Tasks: While medical image segmentation is the primary focus, the architecture is also increasingly applied to classification, image synthesis, reconstruction, and report generation.

3.3 Performance vs. Limitations

Performance: Mamba-based models, particularly hybrids, demonstrate high diagnostic accuracy and often outperform standard CNNs and Transformers in terms of Dice scores and IoU, especially in 3D volumetric tasks where ViTs struggle with memory.
Methodological Bottlenecks:
- Spatial Information Loss: The necessity of flattening 2D/3D images into 1D sequences leads to the loss of spatial continuity and topological consistency.
- Memory Overhead: Multi-directional scanning mechanisms (e.g., in VMamba/SS2D) and hybrid architectures (dual-branch structures) often result in excessive parameter counts and memory consumption, limiting deployment on resource-constrained clinical hardware.
- Generalization: Many studies rely on standard open-access datasets (e.g., BraTS, ISIC) and lack multi-center external validation, raising concerns about robustness against real-world clinical variations (noise, low-dose acquisitions, patient motion).

3.4 Quality Assessment

The review notes that while methodological transparency is generally high (many studies release code), dataset robustness is often low to medium due to a lack of domain generalization. Computational feasibility remains a significant hurdle, with few studies demonstrating successful real-time or edge-device deployment.

4. Significance and Future Directions

The paper posits that Mamba represents a transformative shift in medical image analysis by mathematically reconciling the global context of Transformers with the linear efficiency of CNNs. However, the authors maintain a modest stance, emphasizing that Mamba is not yet a "plug-and-play" solution for clinical integration.

Significance:

Mamba offers a viable alternative to the quadratic bottleneck of ViTs, enabling the processing of high-resolution and 3D medical data without the prohibitive memory costs of Transformers.
The CNN-Mamba hybrid approach is currently the most promising pathway for balancing local and global feature extraction.

Future Directions Identified by the Authors:

Native 3D Selective SSMs: Development of Mamba blocks that process 3D volumes directly without flattening, preserving spatial topology and inter-pixel neighborhood relationships.
Medical Foundation Models: Leveraging Mamba's efficiency to build large-scale, self-supervised foundation models on unlabeled clinical data to address the scarcity of labeled medical datasets.
Edge Integration: Application of model compression techniques (pruning, quantization, knowledge distillation) to reduce parameter density, enabling deployment on portable devices and real-time surgical navigation systems.

The paper concludes that while Mamba-based architectures are rapidly becoming a backbone for next-generation medical AI, overcoming spatial information loss and memory bottlenecks is essential before they can be fully integrated into clinical decision-support systems.