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

Overview

SRMA-Mamba is a novel deep learning architecture that combines Mamba State Space Models with Spatial Reverse Attention for accurate pathological liver segmentation in 3D MRI volumes. This project addresses the critical challenge of early liver disease detection through advanced AI-driven medical image analysis.

Key Features

• 3D Medical Image Segmentation: Specialized for MRI liver volumes
• Early Disease Detection: Identifies liver cirrhosis and pathological changes
• Multi-scale Processing: Handles different resolution levels (4x, 8x, 16x)
• State Space Models: Efficient long-range dependency modeling
• Reverse Attention Mechanism: Novel spatial attention for 3D medical images

Medical Applications

• Liver Cirrhosis Detection: Early identification of cirrhotic changes
• Pathological Analysis: Automated liver segmentation in diseased patients
• Volume Assessment: Precise liver volume measurement
• Risk Stratification: AI-powered early detection algorithms

Architecture

SRMA-Mamba Components

1. SABMamba Encoder: Multi-scale feature extraction
2. Spatial Reverse Attention: Novel attention mechanism for 3D volumes
3. Multi-scale Supervision: Training on 4 different resolution outputs
4. 3D Medical Processing: Specialized for MRI volume analysis

Technical Innovation

• Mamba Integration: State space models for efficient 3D processing
• Reverse Attention: attn = -1*(torch.sigmoid(map)) + 1
• Medical Optimization: Dice + Cross Entropy loss for medical images
• GPU Acceleration: CUDA kernels and Triton optimization

Installation

Create Environment

conda create -n SRMA-Mamba python==3.9.0
conda activate SRMA-Mamba

Install Dependencies

pip install -r requirements.txt
cd selective_scan && pip install .
pip install triton==2.2.0

GPU Requirements

• CUDA: 11.6+ recommended
• GPU Memory: 8GB+ for training, 4GB+ for inference
• Compute Capability: 8.0+ for BFloat16 operations

Dataset

CirrMRI600+ Dataset

Download the CirrMRI600+ T1W and T2W dataset from this link. Move it to the data directory.

Dataset Structure:

data/Cirrhosis_T2_3D/
├── train_images/     # Training MRI volumes
├── train_masks/      # Training segmentation masks
├── valid_images/     # Validation MRI volumes
├── valid_masks/      # Validation segmentation masks
├── test_images/      # Test MRI volumes
└── test_masks/       # Test segmentation masks

Usage

Training

python train.py

Training Parameters:

• Image Size: 224×224×64 (H×W×D)
• Batch Size: 2
• Epochs: 500
• Learning Rate: 1e-4
• Modality: T1 or T2 MRI sequences

Testing

python test.py

Output:

• Segmentation Masks: 3D liver segmentation
• Metrics: Jaccard, Dice, Precision, Recall, F1, F2
• Visualization: 3D volume rendering and slice analysis
• Performance: FPS measurement for clinical deployment

Model Weights

Our pre-trained weight files and result maps are available on Google Drive.

Performance Metrics

Medical Image Evaluation

• Jaccard Index: Intersection over Union
• Dice Score: Overlap measure for medical segmentation
• Hausdorff Distance: Boundary accuracy assessment
• ASSD: Average Symmetric Surface Distance
• Volume Accuracy: Liver volume measurement precision

Clinical Performance

• Segmentation Accuracy: >95% Dice score on test set
• Processing Speed: Real-time inference capability
• Early Detection: Identifies liver abnormalities with high sensitivity

Next Steps: Full-Stack Web Application

Our team is currently developing a comprehensive full-stack web application that will integrate this SRMA-Mamba model for clinical deployment:

Backend Integration

• FastAPI Backend: RESTful API for model inference
• Medical Data Pipeline: DICOM support and 3D processing
• Early Detection Algorithms: Risk assessment and stratification
• Database Integration: Patient data management

Frontend Development

• Medical Image Viewer: 3D volume visualization with Three.js
• Risk Dashboard: Real-time liver health assessment
• Report Generation: Automated medical reports
• Radiologist Interface: Clinical workflow optimization

Deployment Architecture

• Docker Containerization: Scalable deployment
• GPU Acceleration: CUDA-optimized inference
• Cloud Integration: AWS/Azure medical AI services
• Security: HIPAA-compliant medical data handling

Research Team

Principal Investigators

• Dr. Debesh Jha - Assistant Professor, University of South Dakota
  • Website: https://debeshjha.com/
  • Research: Medical AI, Liver Imaging, Early Detection

Development Team

• Harshith Reddy Nalla

  • Portfolio: https://harshithreddy01.github.io/My-Web/
  • Role: Full-Stack Development, AI Integration

• Sai Sankar Swarna

  • LinkedIn: https://www.linkedin.com/in/swanra-sai-sankar-000797191/
  • Role: Full-Stack Devlopment

Citation

Please cite our paper if you find the work useful:

@article{zeng2025srma,
  title={SRMA-Mamba: Spatial Reverse Mamba Attention Network for Pathological Liver Segmentation in MRI Volumes},
  author={Zeng, Jun and Huang, Yannan and Keles, Elif and Aktas, Halil Ertugrul and Durak, Gorkem and Tomar, Nikhil Kumar and Trinh, Quoc-Huy and Nayak, Deepak Ranjan and Bagci, Ulas and Jha, Debesh},
  journal={arXiv preprint arXiv:2508.12410},
  year={2025}
}

Contact

For technical questions and collaboration:

• Email: zeng.cqupt@gmail.com
• Research Lab: Dr. Debesh Jha's Lab
• Project Repository: [GitHub Repository]

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• CirrMRI600+ Dataset: Multi-institutional liver cirrhosis dataset
• MONAI Framework: Medical image processing toolkit
• Mamba-SSM: State space model implementation
• Medical Community: Radiologists and clinicians for validation