Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MambaRecon: MRI Reconstruction with Structured State Space Models

Yilmaz Korkmaz, Vishal M. Patel

TL;DR

This work proposes an inno-vative MRI reconstruction framework that employs structured state space models at its core, aimed at amplifying both long-range contextual sensitivity and reconstruction efficacy.

Abstract

Magnetic Resonance Imaging (MRI) is one of the most important medical imaging modalities as it provides superior resolution of soft tissues, albeit with a notable limitation in scanning speed. The advent of deep learning has catalyzed the development of cutting-edge methods for the expedited reconstruction of MRI scans, utilizing convolutional neural networks and, more recently, vision transformers. Recently proposed structured state space models (e.g., Mamba) have gained some traction due to their efficiency and low computational requirements compared to transformer models. We propose an innovative MRI reconstruction framework that employs structured state space models at its core, aimed at amplifying both long-range contextual sensitivity and reconstruction efficacy. Comprehensive experiments on public brain MRI datasets show that our model sets new benchmarks beating state-of-the-art reconstruction baselines. Code will be available (https://github.com/yilmazkorkmaz1/MambaRecon).

MambaRecon: MRI Reconstruction with Structured State Space Models

TL;DR

This work proposes an inno-vative MRI reconstruction framework that employs structured state space models at its core, aimed at amplifying both long-range contextual sensitivity and reconstruction efficacy.

Abstract

Magnetic Resonance Imaging (MRI) is one of the most important medical imaging modalities as it provides superior resolution of soft tissues, albeit with a notable limitation in scanning speed. The advent of deep learning has catalyzed the development of cutting-edge methods for the expedited reconstruction of MRI scans, utilizing convolutional neural networks and, more recently, vision transformers. Recently proposed structured state space models (e.g., Mamba) have gained some traction due to their efficiency and low computational requirements compared to transformer models. We propose an innovative MRI reconstruction framework that employs structured state space models at its core, aimed at amplifying both long-range contextual sensitivity and reconstruction efficacy. Comprehensive experiments on public brain MRI datasets show that our model sets new benchmarks beating state-of-the-art reconstruction baselines. Code will be available (https://github.com/yilmazkorkmaz1/MambaRecon).
Paper Structure (18 sections, 7 equations, 6 figures, 4 tables)

This paper contains 18 sections, 7 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Data-Driven Convolution Based Models.
  4. Physics-guided Convolution Based Models.
  5. Transformer-based Models.
  6. Mamba in Medical Imaging.
  7. Background
  8. Accelerated MRI Reconstruction
  9. Structured State Space Models
  10. Methodology
  11. Network and Training Details
  12. Experiments
  13. Datasets
  14. Competing Methods
  15. Ablation Studies
  16. ...and 3 more sections

Figures (6)

  • Figure 1: MambaRecon architecture is shown. $\text{X}_{\text{us}}$ corresponds to zero-filled input image, $\text{X}_{\text{fs}}$ is the fully sampled ground truth and $\text{X}_{\text{r}}$ is the reconstructed image. Minimization of $L_1$ norm of the difference between $\text{X}_{\text{fs}}$ and $\text{X}_{\text{r}}$ is utilized as the training objective. Consecutive VSSM and data-consistency blocks (shown as DC Block) are repeated 6 times. $\sum$ corresponds to summation across all unfolded vectors in the output of SSM. DWConv corresponds to the depth-wise convolution howard2017mobilenets.
  • Figure 2: Reconstructions of T2 images with acceleration rate of 4 from fastMRI. Zoomed-in areas and error maps are attached on top and below of reconstructions.
  • Figure 3: Effective receptive fields of each competing method are drawn after training, adapting the codes provided by liu2024vmamba. Gradients are averaged across 100 test slices in fastMRI dataset. For physics-guided methods undersampling mask is given as zero to isolate the effect of the backbone. Each data point in these graphs represents the derivative of the center point in the reconstructed image with respect to the undersampled input image. A higher density of outer points indicates a stronger relationship between the center point and those locations, suggesting greater long-range sensitivity.
  • Figure 4: Reconstructions of Flair images with acceleration rate of 8 from fastMRI. Zoomed-in areas and error maps are attached on top and below of reconstructions.
  • Figure 5: Reconstructions of PD images with acceleration rate of 4 from IXI. Zoomed-in areas and error maps are attached on top and below of reconstructions.
  • ...and 1 more figures