💻 computer science

CS-MUNet: A Channel-Spatial Dual-Stream Mamba Network for Multi-Organ Segmentation

CS-MUNet is a novel dual-stream Mamba network that enhances multi-organ abdominal segmentation by integrating a Boundary-Aware State Mamba module for explicit boundary fusion and a Channel Mamba State Aggregation module for cross-channel semantic collaboration, achieving state-of-the-art performance on public benchmarks.

Original authors: Yuyang Zheng, Mingda Zhang, Jianglong Qin, Qi Mo, Jingdan Pan, Haozhe Hu, Hongyi Huang

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuyang Zheng, Mingda Zhang, Jianglong Qin, Qi Mo, Jingdan Pan, Haozhe Hu, Hongyi Huang

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to solve a giant, messy jigsaw puzzle of a human body, but the pieces are made of soft, squishy clay (like organs) that look almost identical to each other, and the edges where they touch are blurry and hard to see. This is what doctors face when they try to use computers to segment (outline) organs in medical scans like CTs and MRIs.

For a long time, computers were good at finding the "big picture" but terrible at the fine details, like the exact edge of a stomach or a tiny blood vessel. Recently, a new type of AI called Mamba became popular because it's fast and good at looking at long distances. However, the paper argues that existing Mamba models are like a chef who is great at tasting the whole soup but can't tell you exactly where the salt is, or how the different ingredients are mixing together.

The authors of this paper, CS-MUNet, built a new "super-chef" AI that fixes two specific problems using two clever tricks.

The Two Main Problems & The Solutions

1. The "Blurry Edge" Problem

The Issue: In medical scans, the boundary between an organ (like the liver) and the background is often fuzzy. Existing AI models treat the whole image the same way, whether it's the middle of an organ or the edge. It's like trying to paint a picture where you use the same brush pressure for the sky and the fine lines of a bird's feather.

The Solution: The "Boundary-Aware State Mamba" (BASM)
Think of this module as a special pair of glasses that the AI puts on before it starts looking at the image.

How it works: Instead of just guessing where the edge is, this module calculates a "probability map" (a guess of where the edge might be) and feeds that information directly into the AI's brain while it is processing the image.
The Analogy: Imagine you are walking through a dark forest. A normal AI walks with a steady pace everywhere. The CS-MUNet AI, however, has a guide who whispers, "Slow down and pay extra attention here, because you're about to step off a cliff (the edge of an organ)."
The Result: The AI changes how it "forgets" old information and "remembers" new information specifically when it hits a boundary. This makes the outlines of the organs incredibly sharp and precise.

2. The "Teamwork" Problem

The Issue: Medical images have many "channels" (layers of data representing different things like texture, density, or shape). Existing models treat each channel like a solo musician playing in a separate room, never listening to the others. But in reality, the "texture" channel and the "shape" channel need to talk to each other to understand what an organ is.

The Solution: The "Channel Mamba State Aggregation" (CMSA)
Think of this module as a conductor for an orchestra.

How it works: Instead of letting the channels play alone, this module lines them up in a row and forces them to listen to each other in a specific order. It treats the different channels as a sequence of notes in a song.
The Safety Net: The authors realized that if the musicians (channels) talk too freely, they might get confused and create a chaotic noise (mathematically, the data could explode). So, they added a "Lipschitz constraint," which is like a volume limiter or a rulebook. It ensures that while the channels collaborate, they don't get too loud or chaotic.
The Result: The AI understands that a "rough texture" usually goes with a "round shape" for a kidney, but a "smooth texture" goes with a "long shape" for an intestine. This cross-talk helps it distinguish between organs that look similar.

The Big Picture: Why This Matters

The authors tested their new AI on two major public datasets (one with MRI scans of the gut, one with CT scans of the whole abdomen).

The Score: CS-MUNet beat every other top AI model, including the famous ones based on Transformers and older CNNs. It got the highest accuracy scores (mDice) on both tests.
The Efficiency: It didn't just get better; it did so without needing a massive, expensive computer. It's like getting a Ferrari's speed with a Toyota's fuel efficiency.

In Summary

CS-MUNet is a new medical AI that solves the "fuzzy edge" problem by giving the AI special "edge-sensing glasses" (BASM) and solves the "confused data" problem by making the data layers talk to each other like a well-conducted orchestra (CMSA).

By doing this, it can draw much cleaner, more accurate outlines of organs, which helps doctors diagnose diseases faster and more reliably, especially for tricky, hollow organs like the intestines that are hard to see. It's a step forward in making AI a true partner in the operating room.

1. Problem Statement

Medical image segmentation of abdominal hollow organs (e.g., stomach, intestines) is challenging due to variable morphology, indistinct boundaries, and low soft-tissue contrast in CT and MRI scans. While recent Mamba-based (State Space Model) methods have shown promise in medical segmentation by offering linear complexity $O(n)$ for long-range modeling, they suffer from three critical limitations:

Neglect of Cross-Channel Semantic Collaboration: Existing methods treat feature channels as independent parallel components, ignoring the ordered inter-channel dependencies related to anatomical semantics (texture, shape, density).
Superficial Boundary Awareness: Current boundary-aware designs often apply constraints as post-processing or auxiliary losses. They fail to penetrate the core parameter modulation of the State Space Model (SSM) scanning mechanism, meaning boundary information does not actively guide the state transition process.
Lack of Hollow-Organ Optimization: Most existing Mamba methods are validated on solid organs and fail to address the specific thin-wall and weak-boundary challenges of hollow organs.

2. Methodology: CS-MUNet Architecture

The authors propose CS-MUNet, a dual-stream network built on a U-shaped architecture (Res2Net-50 backbone) that introduces two novel modules to address channel and spatial limitations simultaneously.

A. Boundary-Aware State Mamba (BASM)

Location: Deployed at decoder skip connections to fuse shallow encoder features ( $F_e$ ) and deep decoder features ( $F_d$ ).
Core Mechanism: Instead of treating boundary detection as a post-process, BASM embeds boundary awareness directly into the SSM's core parameters ( $\Delta$ and $B$ ).
Bayesian-Attention Framework:
- Generates a pixel-level boundary posterior map ( $P_b$ ) by combining a geometric prior (distance field) and attention likelihood (feature semantics).
- This map is used to decouple into a retain weight ( $R$ ) and an enhance weight ( $E$ ).
Differentiated Parameter Modulation:
- $\Delta$ (Forget Rate): Modulated by $R$ . At boundaries, historical states are retained longer (slower decay) to preserve context.
- $B$ (Write Weight): Modulated by $E$ . At boundaries, the current input weight is amplified to capture fine-grained details.
Dual-Branch Fusion:
- Mamba Branch: Executes the boundary-posterior-driven parameter modulation.
- SASF (Spatial Adaptive Structure-aware Feature) Branch: Uses multi-scale depthwise convolutions (3x3, 5x5, 7x7) to capture local geometric structures, compensating for Mamba's lack of local inductive bias.
- Fusion: The two branches are fused via SE-style weight allocation, ensuring complementary modulation between global sequence modeling and local structural perception.

B. Channel Mamba State Aggregation (CMSA)

Location: Deployed at the encoder bottleneck.
Core Innovation: Redefines the channel dimension as the SSM sequence dimension. Instead of processing channels in parallel, CMSA models them as a sequence to capture inter-channel anatomical dependencies.
Bounded State Transition:
- To prevent unbounded state accumulation across heterogeneous channels (which causes theoretical non-convergence), CMSA introduces Lipschitz constraints.
- It applies clipping to the state transition matrix ( $\bar{A}_k$ ) and the cumulative decay term.
- Channels are grouped ( $G$ groups) to preserve semantic coherence within groups while preventing cross-group contamination.
Result: This enables data-driven modeling of cross-channel semantic collaboration while maintaining stability.

3. Key Contributions

Architectural Paradigm Shift: Proposes a dual-stream SSM modeling pathway that jointly addresses channel semantic collaboration (via CMSA) and boundary-aware feature fusion (via BASM), moving beyond the spatial-only focus of previous Mamba methods.
Deep Boundary Integration: Introduces the BASM module, which uses a Bayesian framework to inject boundary posterior maps directly into the SSM's $\Delta$ and $B$ parameters. This transforms boundary awareness from a passive signal into an active driver of state transitions.
Channel Sequence Modeling: Introduces the CMSA module, which treats channels as a sequence dimension with bounded state transitions, explicitly modeling inter-channel anatomical dependencies that were previously ignored.
State-of-the-Art Performance: Demonstrates consistent superiority over existing CNN, Transformer, and Mamba-based methods on two major benchmarks.

4. Experimental Results

The model was evaluated on two public benchmarks: UW-Madison GI Tract (MRI, hollow organs) and WORD (CT, multi-organ).

Quantitative Performance:
- UW-Madison GI Tract: Achieved 86.16% mDice, outperforming the strongest baseline (MSDUNet) by +2.91%.
- WORD: Achieved 94.47% mDice, outperforming the strongest baseline (Att-UNet) by +1.03%.
- CS-MUNet also achieved superior mIoU scores and lower HD95/ASD (boundary error) metrics compared to all competitors.
Efficiency:
- With 52.2M parameters and 38.8G FLOPs, CS-MUNet is significantly more parameter-efficient than Transformer-based models (e.g., TransUNet: 117M params) while delivering higher accuracy.
Ablation Studies:
- Removing the $\Delta/B$ modulation in BASM caused a drop of ~0.79–1.19% in mDice, proving the necessity of embedding boundary priors into SSM parameters.
- Removing the channel grouping in CMSA degraded performance, confirming the need for bounded inter-channel modeling.

5. Significance and Conclusion

CS-MUNet establishes a new paradigm for State Space Model (SSM) based medical image segmentation. By solving the structural limitations of existing methods—specifically the neglect of cross-channel semantics and the superficial integration of boundary cues—it achieves superior performance in segmenting difficult hollow organs.

The work demonstrates that architectural-level integration of boundary awareness and channel dependencies is more effective than external supervision or post-processing. This makes CS-MUNet a strong candidate for clinical deployment, offering a balance of high segmentation accuracy (critical for hollow organs) and computational efficiency suitable for MRI and CT modalities. Future work aims to extend the framework to 3D volumetric data and further optimize for lightweight deployment.