MCM: Mamba-Based Cardiac Motion Tracking Using Sequential Images in MRI

Jiahui Yin, Xinxing Cheng, Jinming Duan, Yan Pang, Declan P. O'Regan, Hadrien Reynaud, Qingjie Meng

TL;DR

This paper introduces MCM, a novel Mamba-based deep learning framework for tracking cardiac motion in MRI. By leveraging sequential cardiac images and a bi-directional Mamba architecture, MCM achieves smoother and more temporally consistent motion estimations compared to existing methods that rely on single image pairs, leading to improved accuracy in assessing cardiac function.

What Problem Does This Paper Address?

Accurately tracking the motion of the heart's left ventricle (LV) is crucial for assessing cardiac function and diagnosing cardiovascular diseases. Cardiac Magnetic Resonance (CMR) imaging is the gold standard for this, providing detailed sequences of the beating heart. While deep learning has advanced motion tracking, many current methods focus on estimating motion between just two frames (a reference and a single target frame). This approach often results in motion estimates that are inconsistent and lack temporal smoothness, failing to capture the continuous nature of cardiac motion. Incorporating entire image sequences could solve this, but it typically leads to significant memory and computational challenges. This paper aims to bridge this gap by developing a method that effectively utilizes sequential cardiac images for improved motion tracking without prohibitive computational costs.

Methods

The core of the proposed method, MCM, is a Hierarchical Mamba Encoder that processes sequences of target cardiac MRI frames. Instead of just using a single target frame, MCM considers a sequence of frames centered around the target frame ($S_t = \{I_{t-K}, ..., I_t, ..., I_{t+K}\}$). This sequence is paired with the reference end-diastole (ED) frame.

The encoder utilizes Bi-directional Mamba Blocks (BMBs). These blocks are designed to capture spatiotemporal information by employing a novel bi-directional scanning mechanism. This mechanism processes the input sequence in both forward and backward temporal directions, allowing the model to learn more robust and consistent deformation features across different scales.

Following the encoder, a Motion Decoder integrates these multi-scale features. A key component here is the Dual-Path Fusion Head (DFH), which further enhances temporal consistency by processing the fused features in both forward and backward temporal directions before predicting the final motion field ($\Phi_t$). This dual-path approach helps to enforce smoother transitions between motion estimates across consecutive frames.

The model is trained using an end-to-end framework with a combined loss function that includes a similarity loss (measuring how well the warped reference image matches the target image) and a smoothness regularization loss (penalizing jagged motion fields).

Datasets

The paper evaluates MCM on two publicly available cine CMR datasets:

  • ACDC (Automated Cardiac Diagnosis Challenge): This dataset contains short-axis (SAX) slices of the LV.
  • M&Ms (Multi-centre, Multi-vendor and Multi-disease cardiac segmentation): This dataset also provides SAX slices.

Both datasets are preprocessed to a resolution of 1.5 x 1.5 mm, cropped to 128x128 pixels, and normalized to [0, 1]. The datasets are split into training, validation, and testing sets.

Key Insights and Ideas

  • Leveraging Temporal Context is Crucial: The paper strongly demonstrates that considering a sequence of frames, rather than just isolated pairs, significantly improves the temporal consistency and smoothness of cardiac motion tracking.
  • Mamba's Suitability for Sequential Medical Data: The success of Mamba's state-space formulation in handling long sequences is effectively adapted here for the temporal dynamics of cardiac motion, offering an efficient alternative to traditional recurrent or transformer-based approaches for this specific task.
  • Bi-directional Scanning for Enhanced Spatiotemporal Understanding: The proposed bi-directional scanning within the Mamba blocks is a novel contribution that allows for a more comprehensive capture of motion dynamics by considering both past and future temporal information within the sequence.
  • Dual-Path Fusion for Temporal Coherence: The DFH in the motion decoder acts as a crucial refinement step, explicitly enforcing temporal consistency by processing information from both temporal directions, leading to more stable motion field predictions.

Key Results

  • Quantitative Performance: MCM achieved state-of-the-art results on both the ACDC and M&Ms datasets, outperforming conventional methods (dD) and other deep learning approaches (VM, TM, MM) across multiple metrics, including Dice coefficient (accuracy), negative Jacobian determinant (dD), and mean absolute difference of Jacobian determinant (volume preservation). For example, on the ACDC dataset, MCM achieved a Dice score of 83.4(7.1) for the mid-ventricle slice, outperforming the next best method (MM) by 0.7%.
  • Improved Smoothness and Consistency: Qualitative results (Fig. 3 and Fig. 4) show that MCM produces visibly smoother motion fields and more consistent temporal tracking across the cardiac cycle compared to baselines. The method also demonstrates better preservation of volume (lower ||J|-1| values).
  • Computational Efficiency: Despite using sequential images, MCM maintains comparable inference times to baseline methods, with only a moderate increase in GPU VRAM usage proportional to the number of input frames ($N_f$).

Limitations and Future Work

  • 2D Focus: The current work focuses on 2D motion tracking due to the availability of 2D datasets. The authors suggest extending the framework to 3D by integrating 3D convolutions.
  • Hyperparameter Sensitivity: While the paper explores the effect of the smoothness regularization parameter ($\lambda$), it notes that a strong constraint might sacrifice accuracy. Further tuning or adaptive methods could be explored.
  • Dataset Limitations: The evaluation relies on ground truth segmentation being available only for ED and ES frames. The method can estimate motion for all frames, but evaluation is constrained.
  • Future Directions: The authors suggest exploring the framework for 3D motion tracking and potentially integrating it with other imaging modalities or tasks.

Summary updated

BibTeX
@inproceedings{yin2025mcm,
  author = {Jiahui Yin and Xinxing Cheng and Jinming Duan and Yan Pang and Declan P. O'Regan and Hadrien Reynaud and Qingjie Meng},
  title = {{MCM: Mamba-Based Cardiac Motion Tracking Using Sequential Images in MRI}},
  booktitle = {GRAIL/RIME @ MICCAI},
  year = {2025},
  doi = {10.1007/978-3-032-06103-4\_5},
}

Loading related content...

Figure