SVRMamba: Slice-to-Volume Reconstruction from Multiple MRI Stacks with Slice Sequence Guided Mamba

Jiangjie Wu, Hongjiang Wei, Yuyao ZhangAAAI 2025 · DOI

TL;DR

This paper introduces SVRMamba, a novel framework for reconstructing high-resolution 3D fetal brain MRIs from multiple 2D slice stacks. By adapting the Mamba architecture for slice sequence modeling and integrating a CNN-based interpolation network, SVRMamba significantly accelerates reconstruction speed (60x) while achieving state-of-the-art accuracy, addressing the critical need for faster and more robust fetal MRI analysis.

What Problem Does This Paper Address?

Fetal magnetic resonance imaging (MRI) is crucial for studying brain development, especially in cases of severe fetal motion. However, acquiring 3D fetal MRI data is challenging. The standard approach involves acquiring multiple 2D slices and then reconstructing them into a 3D volume, a process known as slice-to-volume reconstruction (SVR). This SVR process is plagued by issues like slice misalignment and image noise, making it computationally intensive and prone to errors. Existing "coarse-to-fine" methods, which iteratively correct motion and reconstruct the volume, are slow and struggle with accurately modeling the complex spatial relationships between slices, leading to suboptimal results. This paper aims to overcome these limitations by developing a faster and more accurate SVR method that better captures inter-slice spatial dependencies.

Methods

SVRMamba introduces two key innovations:

  1. Slice Sequence-Guided State Space Modeling: The core of SVRMamba is a novel adaptation of the Mamba architecture, a powerful sequence modeling tool. Instead of Mamba's standard unidirectional scanning, SVRMamba employs a "slice sequence-guided odd-even directional scanning" approach. This allows it to effectively capture spatial relationships not just within a single slice sequence but also between different stacks of slices. To further enhance this, they introduce "slice sequence embedding tokens" that mark the position of each slice within its stack and the overall stack sequence. This helps the model understand the context of each slice.
  2. CNN-Based Interpolation Network: After predicting motion parameters for each slice using the Mamba-based network, the slices are stacked to form an initial 3D volume. This volume often contains holes due to missing data. To address this, SVRMamba integrates a fully convolutional neural network (CNN) interpolation network. This network effectively fills these gaps and, importantly, suppresses noise, leading to a cleaner and more complete 3D reconstruction.

Datasets

The paper utilizes three publicly available datasets for training and evaluation:

  • FeTA dataset: Contains 80 fetal brain MRI volumes (68 for training, 12 for testing).
  • FBA dataset: Features 14 fetal brain MRI volumes with 0.8 mm isotropic resolution (6 for training, 8 for testing).
  • Spina Bifida Aperta (SBA) dataset: Includes 14 pathological fetal brain MRI volumes with 0.8 mm isotropic resolution (6 for training, 8 for testing).

For training, they created paired data by simulating three-view orthogonal stacks, motion parameters, and ground truth volumes from registered GT volumes, introducing random transformations to mimic motion.

Key Insights and Ideas

  • Mamba for Spatial Sequences: The paper demonstrates that Mamba, typically used for temporal sequences, can be effectively adapted for modeling spatial relationships between MRI slices, a key challenge in SVR.
  • Odd-Even Scanning for Slices: Reformulating Mamba's scanning to be "slice sequence-guided odd-even" is crucial for capturing the interleaved nature of fetal MRI acquisition and the relationships between odd and even slices.
  • Embedding for Context: Using slice sequence embedding tokens to mark slice and stack positions provides essential contextual information for the state space model.
  • CNN Interpolation for Noise Reduction: The integrated CNN interpolation network is not just for filling holes but also plays a vital role in denoising the final 3D volume, a significant improvement over standard interpolation methods.
  • Speed-Accuracy Trade-off: The proposed method achieves a remarkable speedup (60x) without sacrificing, and in fact improving, reconstruction accuracy.

Key Results

  • State-of-the-Art Performance: SVRMamba outperforms existing state-of-the-art (SOTA) methods on benchmark and clinical datasets across various metrics, including PSNR, SSIM, MT (translation error), and MR (rotation error).
  • Significant Speedup: The reconstruction time is reduced by approximately 60 times compared to other methods. For example, SVRMamba takes only 3.0 seconds, while other methods range from 183.9 to 1642.8 seconds.
  • Improved Reconstruction Quality: Qualitative comparisons show SVRMamba produces cleaner, more detailed, and less noisy 3D reconstructions, particularly evident on datasets with significant motion and noise (e.g., SBA dataset).
  • Ablation Study Findings: Ablation experiments confirm the critical importance of the Slice Sequence-guided State Space Network (SSC) and Slice Sequence Embedding (SSE) for capturing local context and spatial sequences. The CNN-based Interpolation Network (IN) is also shown to be essential for high-quality reconstruction.

Limitations and Future Work

  • Generalization to Real Clinical Data: While the paper tests on real clinical datasets, it notes that the primary training was on simulated data. Further validation on larger, more diverse real clinical datasets would be beneficial.
  • Interpolation Network Focus: The authors state that searching for the "best interpolation network model is not the focus of this study," implying that further optimization of the IN could yield even better results.
  • Future Directions: The paper suggests that the framework's ability to capture spatial relationships and its speed make it suitable for real-time applications and further exploration in other medical imaging reconstruction tasks. The success of Mamba for spatial sequences opens avenues for its application in other image processing domains.

Summary updated

BibTeX
@inproceedings{wu2025svrmamba,
  author = {Jiangjie Wu and Hongjiang Wei and Yuyao Zhang},
  title = {{SVRMamba: Slice-to-Volume Reconstruction from Multiple MRI Stacks with Slice Sequence Guided Mamba}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i8.32907},
}
Figure