Meta-learning Slice-to-Volume Reconstruction in Fetal Brain MRI using Implicit Neural Representations

TL;DR

The paper tackles the challenging problem of reconstructing a high-resolution fetal brain volume $V$ from motion-corrupted 2D slices using a fully implicit neural representation (INR) framework. It introduces two sine-activated MLPs (Slice Module and SR Module) that jointly perform motion correction, outlier handling, and super-resolution, connected through a forward model $X_i = B_i T_i V + \epsilon_i$ and optimized with a robust mean absolute error via Monte Carlo PSF sampling. A key advance is self-supervised meta-learning, yielding two meta-initializations that enable rapid adaptation across reconstruction tasks and centers, with reported speedups of up to 50% and improved robustness under severe artifact and motion corruption. The approach is validated on simulated data and real fetal MRI from dHCP and CHUV (over 480 reconstructions across >160 subjects), showing superior reconstruction quality under challenging conditions and demonstrating practical viability for clinical deployment with limited training data.

Abstract

High-resolution slice-to-volume reconstruction (SVR) from multiple motion-corrupted low-resolution 2D slices constitutes a critical step in image-based diagnostics of moving subjects, such as fetal brain Magnetic Resonance Imaging (MRI). Existing solutions struggle with image artifacts and severe subject motion or require slice pre-alignment to achieve satisfying reconstruction performance. We propose a novel SVR method to enable fast and accurate MRI reconstruction even in cases of severe image and motion corruption. Our approach performs motion correction, outlier handling, and super-resolution reconstruction with all operations being entirely based on implicit neural representations. The model can be initialized with task-specific priors through fully self-supervised meta-learning on either simulated or real-world data. In extensive experiments including over 480 reconstructions of simulated and clinical MRI brain data from different centers, we prove the utility of our method in cases of severe subject motion and image artifacts. Our results demonstrate improvements in reconstruction quality, especially in the presence of severe motion, compared to state-of-the-art methods, and up to 50% reduction in reconstruction time.

Figures (6)

• Figure 1: Overview 3D reconstruction. a) Acquisition stacks of motion corrupted 2D slices across different anatomical planes. b) Architecture: Our framework comprises two key components: the Slice Module and the SR Module . The Slice Module operates on positional encodings of slices, termed slice encodings, and is designed with two output heads. One head estimates motion correction while the other head addresses outlier handling, including slice intensity scaling ($\sigma$) and slice weighting ($\omega$). Simultaneously, the SR Module learns a continuous super-resolved and motion-free 3D brain representation by predicting voxel intensities of coordinates sampled from the motion corrected slices. c) After optimization, we obtain the high-resolution 3D reconstruction by querying the SR Module with coordinates sampled from a 3D grid.
• Figure 2: Real-world raw acquisition stacks from dHCP (top) and CHUV (bottom). Figure created with ITK-SNAP itksnap
• Figure 3: Evaluation of five simulated datasets with increasing simulated motion- and image corruption from left $(\mu=1)$ to right $(\mu=5)$. 2D visualization of 3D data. a) Reconstruction results of five cases with increasing simulated corruption. Top row shows one of three acquisition stacks used as input for the reconstruction process. Subsequent rows show reconstruction results of the baselines and our proposed method with standard initialization (SSVR) and meta-learned initialization ($\text{SSVR}_{\text{meta1}})$. b) Depicted are mean PSNR (solid lines) and SSIM (dashed lines) for the five simulated datasets with 20 cases each. X-axis represents the corruption factor applied to the dataset with $\mu=1$ indicating minimal corruption and $\mu=5$ extreme corruption (see Section \ref{['dHCP_synthetic_data']} for more details).
• Figure 4: Reconstruction results of four cases of different weeks GA from the dHCP dataset. Columns show the results of each method. Baselines struggle particularly with younger cases which typically exhibit stronger motion corruption. Our method consistently produces high-quality reconstructions with standard (SSVR) and meta-learned initialization ($\text{SSVR}_{\text{meta2}}$).
• Figure 5: Stack ablation on 40 fetal subjects of the dHCP dataset. a) Leftmost column shows the reference generated as described in Section \ref{['subsec:evaluation_refs']}. Remaining columns present the reconstruction results for each method (rows) across the four stack ablation scenarios ranging from six, i.e., all available acquisition stacks, down to three stacks used for reconstruction. b) Mean PSNR (solid lines) and SSIM (dashed lines) of all methods for the different stack ablations counting 40 cases each.