🤖 AI

SRMA-Mamba: Spatial Reverse Mamba Attention Network for Pathological Liver Segmentation in MRI Volumes

The paper introduces SRMA-Mamba, a novel Mamba-based network that leverages a Spatial Anatomy-Based Mamba module and a Spatial Reverse Mamba Attention module to effectively model complex spatial relationships and refine boundary details for state-of-the-art 3D pathological liver segmentation in MRI volumes.

Original authors: Jun Zeng, Quoc-Huy Trinh, Deepak Ranjan Nayak, Nikhil Kumar Tomar, Ulas Bagci, Debesh Jha

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jun Zeng, Quoc-Huy Trinh, Deepak Ranjan Nayak, Nikhil Kumar Tomar, Ulas Bagci, Debesh Jha

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

The Big Problem: Finding the "Bad" Liver in a 3D Puzzle

Imagine a patient has a liver that is sick (cirrhotic). Doctors need to see exactly where the sick parts are to help the patient. Usually, they look at 3D MRI scans, which are like thousands of thin slices of bread stacked together to make a loaf.

Currently, doctors have to look at these slices one by one and draw the outline of the sick liver by hand. This is:

Tiring: It takes a long time.
Inconsistent: One doctor might draw the line slightly differently than another.
Hard: The sick liver often looks messy and irregular, making it hard to tell where it starts and stops.

Computer programs (AI) exist to do this automatically, but many of them struggle. They often treat the 3D liver like a long, flat list of numbers (like unrolling a rug), which causes them to lose the "3D shape" and the spatial relationships between different parts of the liver. This leads to messy, inaccurate outlines.

The Solution: SRMA-Mamba

The authors created a new AI system called SRMA-Mamba. Think of it as a super-smart robot assistant designed specifically to look at 3D liver scans and draw perfect outlines of the sick parts.

Here is how it works, broken down into three main tricks:

1. The "Three-Angle Detective" (SABMamba)

Most AI looks at a 3D object from just one direction or flattens it out. But imagine you are trying to understand a complex sculpture. If you only look at it from the front, you miss the sides and the back.

How SRMA-Mamba does it: It acts like a detective who walks around the liver looking at it from three specific angles at once: Front-to-Back (Axial), Side-to-Side (Coronal), and Top-to-Bottom (Sagittal).
The Analogy: Instead of looking at a single slice of bread, this AI looks at the whole loaf from the top, the side, and the end simultaneously. It combines these three views to understand the true 3D shape of the sick liver, ensuring it doesn't miss any hidden corners.

2. The "Smart Filter" (ABSS)

Once the AI is looking at the liver from all three angles, it has to decide what information is important and what is just background noise.

How SRMA-Mamba does it: It uses a "Selective Scan" mechanism.
The Analogy: Imagine you are in a crowded room trying to hear one specific person's voice. You don't listen to everyone equally; you tune your ears to focus only on that one voice and ignore the chatter. This module does the same thing with the MRI data: it filters out the "noise" (healthy tissue or empty space) and focuses its attention only on the "signal" (the pathological liver tissue).

3. The "Fine-Tuner" (SRMA)

After the AI makes a first guess at where the sick liver is, the outline might be a little rough or blurry, like a sketch drawn with a thick marker.

How SRMA-Mamba does it: It uses a "Reverse Attention" module.
The Analogy: Imagine an artist who draws a rough sketch of a face, then steps back, looks at the mistakes, and uses a fine-tipped pen to clean up the edges and fix the details. This module takes the "rough sketch" (the initial guess) and the "features" (the details it found earlier) and works backward to sharpen the edges. It specifically targets the boundaries to make sure the line between "sick liver" and "healthy liver" is crisp and accurate.

What Did They Find?

The authors tested this new robot assistant on a large collection of real liver MRI scans (628 scans total). They compared it against other top-tier AI systems.

The Results: SRMA-Mamba won. It was more accurate than the other methods at finding the sick liver.
The Metrics: It got higher scores for how well it matched the "truth" (Dice score) and had fewer errors in the distance of the edges (Hausdorff distance).
Efficiency: It also did this while using less computer power than some of the other heavy-duty models.

Summary

In short, SRMA-Mamba is a new AI tool that looks at 3D liver scans from multiple angles at once, filters out the noise to focus on the disease, and then refines its drawing to get the edges perfect. It proved to be the best at automatically finding and outlining sick livers in MRI scans compared to current technology.

Technical Summary: SRMA-Mamba for Pathological Liver Segmentation

Problem Statement

Liver cirrhosis represents a critical global health challenge, serving as the end-stage manifestation of progressive liver fibrosis. While Magnetic Resonance Imaging (MRI) has become a cornerstone for non-invasive diagnosis due to its superior soft-tissue contrast, the accurate delineation of pathological liver structures remains difficult. Manual segmentation is labor-intensive and suffers from significant inter-observer variability.

Existing automated methods, particularly those based on deep learning, face specific limitations in this domain:

Underutilization of Spatial Context: Many approaches treat volumetric MRI data as linear sequences, which inherently disrupts the intrinsic 3D spatial topology.
Anisotropic Morphologies: Cirrhotic structures exhibit highly irregular and anisotropic shapes. Discarding spatial context in favor of linearization leads to suboptimal boundary definitions and fragmented segmentation results.
Complex Anatomical Relationships: Current methods often fail to fully capture the intricate spatial relationships between voxels across different anatomical planes, hindering clinical explainability and effectiveness.

Methodology

The authors propose SRMA-Mamba (Spatial Reverse Mamba Attention Network), a novel architecture designed to model spatial relationships within complex anatomical structures of MRI volumes using State-Space Models (SSMs). The network follows an encoder-decoder structure with three core innovations:

1. Overall Architecture

SRMA-Mamba employs a hierarchical encoder-decoder framework. The encoder extracts features from input cirrhotic MRI volumes, while the decoder incrementally refines segmentation maps. Unlike standard decoders, this architecture integrates a Spatial Reverse Mamba Attention (SRMA) module at multiple stages to progressively refine edge details using features from both the encoding stages and coarse segmentation maps.

2. Spatial Anatomy-Based Mamba (SABMamba)

Inspired by the clinical practice of radiologists reviewing MRI volumes across multiple planes, the SABMamba module models complex anatomical relationships by integrating information from three orthogonal planes: sagittal, coronal, and axial.

Structure: It consists of consecutive residual branches. One branch utilizes an Anatomy-Based Mamba (ABM) block, while the other uses a Multi-Layer Perceptron (MLP).
Anatomy-Based Selective Scan (ABSS): A key component within the ABM block, the ABSS module learns spatial feature representations for pathological liver tissues. It operates by:
1. Reshaping: Rearranging the input volume into three anatomical planes ( $f_h, f_w, f_d$ ).
2. Scan Expanding: Unfolding data from these planes into sequences along four distinct directions.
3. Selective Processing: Utilizing the S6 module (from Mamba) to selectively process these sequences, filtering irrelevant information while preserving relevant features.
4. Scan Merging: Integrating the processed sequences back into the original resolution.
  This approach allows for direct processing of 3D volumetric data with low computational complexity while preserving 3D spatial topology.

3. Spatial Reverse Mamba Attention (SRMA)

The SRMA module is designed to progressively refine boundary details in the segmentation map.

Mechanism: It takes the coarse segmentation map ( $S_{i+1}$ ) and the feature map from the current encoding level ( $f_i$ ) as inputs.
Reverse Attention: The coarse segmentation map is passed through a Sigmoid function and a reverse operation to generate attention weights ( $A_i = 1 - \text{Sigmoid}(S_{i+1})$ ).
Refinement: These weights are applied via element-wise multiplication to the enhanced feature map. The result is processed by a SABMamba module and convolutional layers to generate the refined output. This process enhances feature aggregation and boundary delineation.

Key Contributions

The paper outlines four primary contributions:

SRMA-Mamba Network: A novel network for efficient volumetric pathological liver segmentation that addresses the limitations of linear sequence modeling in 3D medical imaging.
SABMamba Module: A mechanism that performs scans across multiple anatomical perspectives (sagittal, coronal, axial) to construct a global spatial context representation.
SRMA Module: A progressive refinement mechanism that utilizes hierarchical encoding features and coarse segmentation maps to improve boundary precision.
ABSS Module: An Anatomy-Based Selective Scan module that enables the direct processing of 3D volumetric data with low computational complexity, preserving spatial anatomical information.

Experimental Results

The authors evaluated SRMA-Mamba on the CirrMRI600+ dataset, which includes 628 abdominal MRI scans (310 T1-weighted and 318 T2-weighted). The model was compared against state-of-the-art methods including CNNs (SegResNet, UX-Net, MedNeXt), Transformers (SwinUNETR, SwinUNETRv2), and Mamba-based models (SegMamba).

Quantitative Performance

T1W Dataset: SRMA-Mamba achieved a 92.95% mDSC and 87.59% mIoU, outperforming the previous best (SegMamba) by 1.15% and 1.89%, respectively. It also demonstrated superior boundary accuracy with an HD95 of 17.52 and ASSD of 3.05.
T2W Dataset: The model achieved an 86.25% mDSC and 78.24% mIoU, again surpassing SegMamba.
Computational Efficiency: Despite high performance, SRMA-Mamba demonstrated significantly lower computational complexity compared to SegMamba, with 17.22M parameters and 149.14 GFLOPs, compared to SegMamba's 64.24M parameters and 2379.03 GFLOPs.

Qualitative and Ablation Analysis

Qualitative Results: Visual comparisons showed SRMA-Mamba's superior capability in delineating both large and small pathological regions and handling complex boundary details compared to other methods.
Ablation Study: Removing SABMamba reduced mDSC from 92.95% to 80.62%, confirming its role in modeling spatial anatomy. Removing SRMA reduced mDSC to 90.63%, validating its contribution to boundary refinement.
Plane Configuration: Experiments showed that using all three anatomical planes (Axial, Sagittal, Coronal) consistently outperformed single-plane configurations, validating the multi-plane representation learning approach.

Significance and Claims

The paper claims that SRMA-Mamba addresses the critical gap in existing methods by effectively modeling the 3D spatial topology of pathological liver structures without the fragmentation caused by linearization. By integrating anatomical plane information and employing a reverse attention mechanism, the network achieves state-of-the-art performance in 3D pathological liver segmentation. The authors assert that this approach not only improves segmentation accuracy but also enhances clinical explainability by better preserving the intricate anatomical details required for prognosis and intervention strategies. The work positions the integration of State-Space Models with anatomical priors as a promising direction for volumetric medical image analysis.