MoRe-3DGSMR: Motion-resolved reconstruction framework for free-breathing pulmonary MRI based on 3D Gaussian representation

TL;DR

The paper tackles motion-induced degradation in high-resolution free-breathing pulmonary MRI by introducing MoRe-3DGSMR, an unsupervised framework that fuses a 3D Gaussian representation (3DGS) with a DVF-based motion decoder. Respiratory motion is inferred from center-k-space data to bin radial acquisitions into motion states; the first state is reconstructed with 3DGS, and a CNN predicts deformation fields to generate remaining states, all optimized end-to-end with data fidelity (NUFFT) and TV regularization. A novel multi-resolution initialization initializes Gaussian points across frequency bands, enabling simultaneous multi-scale training and faster convergence without adaptive densification. Experimental results across six subjects show MoRe-3DGSMR achieves higher SNR and CNR than XD-GRASP, ExtremeMRI, and Moco-SToRM while preserving fine anatomical detail, highlighting its potential for clinical 3D isotropic pulmonary MRI under free-breathing conditions.

Abstract

This study presents an unsupervised, motion-resolved reconstruction framework for high-resolution, free-breathing pulmonary magnetic resonance imaging (MRI), utilizing a three-dimensional Gaussian representation (3DGS). The proposed method leverages 3DGS to address the challenges of motion-resolved 3D isotropic pulmonary MRI reconstruction by enabling data smoothing between voxels for continuous spatial representation. Pulmonary MRI data acquisition is performed using a golden-angle radial sampling trajectory, with respiratory motion signals extracted from the center of k-space in each radial spoke. Based on the estimated motion signal, the k-space data is sorted into multiple respiratory phases. A 3DGS framework is then applied to reconstruct a reference image volume from the first motion state. Subsequently, a patient-specific convolutional neural network is trained to estimate the deformation vector fields (DVFs), which are used to generate the remaining motion states through spatial transformation of the reference volume. The proposed reconstruction pipeline is evaluated on six datasets from six subjects and bench-marked against three state-of-the-art reconstruction methods. The experimental findings demonstrate that the proposed reconstruction framework effectively reconstructs high-resolution, motion-resolved pulmonary MR images. Compared with existing approaches, it achieves superior image quality, reflected by higher signal-to-noise ratio and contrast-to-noise ratio. The proposed unsupervised 3DGS-based reconstruction method enables accurate motion-resolved pulmonary MRI with isotropic spatial resolution. Its superior performance in image quality metrics over state-of-the-art methods highlights its potential as a robust solution for clinical pulmonary MR imaging.

Figures (12)

• Figure 1: Illustration of the multi-resolution initialization strategy. Gaussian points are initialized at multiple spatial scales using outputs from inverse Fourier transforms computed at progressively increasing levels of k-space coverage, forming a hierarchical (pyramidal) architecture. Each resolution level corresponds to a subset of the k-space data, with higher levels incorporating a greater number of samples per radial spoke to progressively recover higher-frequency information. The Gaussian points initialized at each level are summed, and then simultaneously refined during the training process.
• Figure 2: Overall pipeline for MoRe-3DGSMR. The framework comprises the following key components: k-space binning, multi-resolution Gaussian points initialization, voxelization, DVFs prediction, motion warping, and NUFFT operations. The estimated respiratory signal is used to retrospectively sort radial k-space spokes into discrete motion states corresponding to different respiratory phases. The k-space data from the first motion state are utilized for initializing the Gaussian points. During training, these Gaussian points are iteratively optimized and subsequently voxelized into volumetric representations.
• Figure 3: Examples of predicted DVFs. (a) Coronal view. (b) Sagittal view.
• Figure 4: Results of the ablation study investigating the impact of the number of initialized Gaussian points and the number of training iterations are presented. Reconstruction performance was quantitatively evaluated using four metrics -- Sobel index, Laplacian index, SNR, and CNR -- across four anatomical regions: lung, liver, heart, and spine.
• Figure 5: Results of the ablation study evaluating the effect of the hyperparameter $\lambda_{tv}$ are presented. Quantitative assessments were performed using four evaluation metrics (a) Sobel index, (b) Laplacian index, (c) SNR, and (d) CNR -- across four anatomical regions: lung, liver, heart, and spine.