Dual Correlation-aware Mamba for Microvascular Obstruction Identification in Non-contrast Cine Cardiac Magnetic Resonance

September 23, 2025 · View on GitHub

This repository contains the implementation of Dual Correlation-aware Mamba (DCMamba) for microvascular obstruction identification in non-contrast cine cardiac magnetic resonance imaging.

Overview

DCMamba is a novel architecture that combines Mamba state-space models with dual correlation mechanisms for efficient microvascular obstruction (MVO) identification in non-contrast cine cardiac MR sequences. The model leverages both spatial and temporal correlations to detect subtle MVO patterns without requiring contrast enhancement.

Environment Setup

Requirements

  • Python >= 3.8
  • PyTorch >= 1.8
  • CUDA 11.8
  • MMSegmentation framework
  • mamba-ssm

Installation

  1. Clone the repository:
git clone https://github.com/code-koukai/Dual-Correlation-Mamba.git
cd Dual-Correlation-Mamba
  1. Install dependencies:
pip install torch torchvision torchaudio
pip install mmengine mmcv
pip install mamba-ssm
pip install timm
  1. Install MMSegmentation:
pip install -e .

Project Structure

MICCAI2025/
├── configs/                    # Configuration files
│   ├── DCMamba.py             # Main model configuration
│   └── _base_/                # Base configurations
├── mmseg/                     # MMSegmentation framework
│   ├── models/
│   │   ├── backbones/
│   │   │   └── deformable_ablation/
│   │   │       └── dcmamba.py # DCMamba backbone implementation
│   │   └── decode_heads/      # Segmentation heads
│   └── datasets/              # Dataset implementations
├── tools/                     # Training and testing scripts
│   ├── train.py              # Training script
│   └── test.py               # Testing script
└── README.md                 # This file

Usage

Training

To train the DCMamba model:

python tools/train.py configs/DCMamba.py --work-dir ./work_dirs/dcmamba_experiment

Training Options:

  • --work-dir: Directory to save logs and model checkpoints
  • --resume: Resume training from the latest checkpoint
  • --amp: Enable automatic mixed precision training
  • --cfg-options: Override config options

Example with additional options:

python tools/train.py configs/DCMamba.py \
    --work-dir ./work_dirs/dcmamba_experiment \
    --amp \
    --cfg-options train_dataloader.batch_size=8

Testing/Evaluation

To evaluate a trained model:

python tools/test.py configs/DCMamba.py /path/to/checkpoint.pth

Testing Options:

  • --work-dir: Directory to save evaluation results

Example:

python tools/test.py configs/DCMamba.py \
    ./work_dirs/dcmamba_experiment/iter_10000.pth \
    --work-dir ./work_dirs/dcmamba_test \

Model Configuration

The main configuration file configs/DCMamba.py contains:

  • Model Architecture: DCMamba backbone with BertPseudoHead
  • Loss Functions: CrossEntropyLoss + MAEloss
  • Optimizer: AdamW with different learning rates for different components
  • Training Settings: 10k iterations with adaptive learning rate

Update the dataset configuration in configs/_base_/datasets/dataset.py

Acknowledgments

  • MMSegmentation framework
  • Mamba state-space models
  • TIMM library for vision transformers