SUFFICIENT: A scan-specific unsupervised deep learning framework for high-resolution 3D isotropic fetal brain MRI reconstruction

TL;DR

SUFFICIENT addresses the challenge of high-resolution isotropic 3D fetal brain MRI reconstruction from motion-corrupted 2D slices without reliance on external training data. It jointly optimizes a scan-specific unsupervised SVR-net and an unsupervised SRR-net, using a differentiable forward model and a DIP-inspired decoding approach to enforce local consistency and robustness to artifacts. The method demonstrates superior performance over four state-of-the-art baselines on simulated and clinical fetal MRI data, with strong improvements in PSNR, SSIM, and NCC, and robust motion correction evidenced by lower MAE/RMSE in estimated Transform parameters. This framework offers a practical, data-efficient pathway toward reliable clinical 3D fetal brain reconstructions across varying slice thicknesses and imaging conditions.

Abstract

High-quality 3D fetal brain MRI reconstruction from motion-corrupted 2D slices is crucial for clinical diagnosis. Reliable slice-to-volume registration (SVR)-based motion correction and super-resolution reconstruction (SRR) methods are essential. Deep learning (DL) has demonstrated potential in enhancing SVR and SRR when compared to conventional methods. However, it requires large-scale external training datasets, which are difficult to obtain for clinical fetal MRI. To address this issue, we propose an unsupervised iterative SVR-SRR framework for isotropic HR volume reconstruction. Specifically, SVR is formulated as a function mapping a 2D slice and a 3D target volume to a rigid transformation matrix, which aligns the slice to the underlying location in the target volume. The function is parameterized by a convolutional neural network, which is trained by minimizing the difference between the volume slicing at the predicted position and the input slice. In SRR, a decoding network embedded within a deep image prior framework is incorporated with a comprehensive image degradation model to produce the high-resolution (HR) volume. The deep image prior framework offers a local consistency prior to guide the reconstruction of HR volumes. By performing a forward degradation model, the HR volume is optimized by minimizing loss between predicted slices and the observed slices. Comprehensive experiments conducted on large-magnitude motion-corrupted simulation data and clinical data demonstrate the superior performance of the proposed framework over state-of-the-art fetal brain reconstruction frameworks.

Figures (7)

• Figure 1: Our framework for fetal brain MRI reconstruction aims to achieve HR fetal brain volume from multiple motion-corrupted input LR stacks. The framework comprises an alternatingly optimized SVR-net and an SRR-net. During training, the SVR-net initially aligns the motion-corrupted fetal brain stacks to a canonical 3D space and transfers the motion-corrected matrices to the SRR-net. The SRR-net, receiving the original stack of slices and these matrices, generates the 3D fetal brain volume from the motion-corrected MR slices. This reconstructed volume is then fed back to the SVR-net for further refinement of registration accuracy. Both networks undergo alternating optimization until convergence is achieved for the final high-resolution (HR) fetal brain output. In the inference stage, the high-resolution output from the SRR-net represents the final HR fetal brain image. Details about SVR and SRR nets are presented in details in sections \ref{['subsec:svrnet']} and \ref{['subsec:srrnet']}, respectively.
• Figure 2: The proposed unsupervised SVR network for fetal brain motion correction. It predicts the transform matrix $\boldsymbol{t}_k$ that aligns the input slice $\mathbf{y}_k$ to 3D volume $\mathbf{x}$. The SVR-net comprises a dual-branch localization-net, and a differentiable slice acquisition model. During training, the network minimizes the error between the input slice $\mathbf{y}_k$ and the estimated slice $\hat{\mathbf{y}}_k$, which is computed using the slice acquisition model $\hat{\mathbf{y}}_k=$$C_k\mathbf{A}\left(T\left(\mathbf{x}, \boldsymbol{t}_k\right) ; \Sigma_k\right)$ (E.q \ref{['equation:forward']}).
• Figure 3: The proposed unsupervised SRR network for $3 \mathrm{D}$ volume reconstruction. It consists of a decoding network and a slice generation process. The decoding network takes a fixed noise $\mathbf{z}$ as input and maps it to an HR volume $\mathbf{x}$, which represents the fine fetal brain reconstruction from input motion-corrected brain slices. The motion correction for input slices is conducted using transformation matrix $t$ predicted by SVR-net. The SRR network is optimized by minimizing the difference between the predicted slices $\left\{\hat{\mathbf{y}}_k\right\}_{k=1}^N$ and the input slices $\left\{\mathbf{y}_k\right\}_{k=1}^N$, where $\hat{\mathbf{y}}_k=C_k\mathbf{A}\left(T\left(\mathbf{x}, \boldsymbol{t}_k\right) ; \Sigma_k\right)$ computed from the slice generation model in section \ref{['subsec:forwardmodel']}.
• Figure 4: A representative simulated stack sample from Group B in coronal (left), axial (mid) and sagittal (right) views.
• Figure 5: (a) The qualitative reconstruction results obtained on the Group A dataset of a representative subject at a 27-week GA. (b) The qualitative reconstruction results obtained on the Group B dataset of a representative subject at a 23-week GA.