Missing Data Estimation for MR Spectroscopic Imaging via Mask-Free Deep Learning Methods

TL;DR

This work tackles the challenge of missing or corrupted MRSI data in high-resolution acquisitions by introducing a mask-free deep learning framework that uses 2D and 3D U-Nets. A synthetic data pipeline generates HR MRSI volumes from multimodal MRI and simulates localized voxel loss, training with a progressive degradation strategy to enhance robustness. The approach achieves strong quantitative performance ($MSE$ ≈ $0.002$ and $SSIM$ ≈ $0.97$ for 2D; $MSE$ ≈ $0.001$ and $SSIM$ ≈ $0.98$ for 3D) and demonstrates good generalization to real datasets without retraining. This method offers a practical path toward reliable, high-fidelity restoration of metabolite maps in clinical and research settings.

Abstract

Magnetic Resonance Spectroscopic Imaging (MRSI) is a powerful tool for non-invasive mapping of brain metabolites, providing critical insights into neurological conditions. However, its utility is often limited by missing or corrupted data due to motion artifacts, magnetic field inhomogeneities, or failed spectral fitting-especially in high resolution 3D acquisitions. To address this, we propose the first deep learning-based, mask-free framework for estimating missing data in MRSI metabolic maps. Unlike conventional restoration methods that rely on explicit masks to identify missing regions, our approach implicitly detects and estimates these areas using contextual spatial features through 2D and 3D U-Net architectures. We also introduce a progressive training strategy to enhance robustness under varying levels of data degradation. Our method is evaluated on both simulated and real patient datasets and consistently outperforms traditional interpolation techniques such as cubic and linear interpolation. The 2D model achieves an MSE of 0.002 and an SSIM of 0.97 with 20% missing voxels, while the 3D model reaches an MSE of 0.001 and an SSIM of 0.98 with 15% missing voxels. Qualitative results show improved fidelity in estimating missing data, particularly in metabolically heterogeneous regions and ventricular regions. Importantly, our model generalizes well to real-world datasets without requiring retraining or mask input. These findings demonstrate the effectiveness and broad applicability of mask-free deep learning for MRSI restoration, with strong potential for clinical and research integration.

Figures (7)

• Figure 1: Visualization of 3D MRSI metabolic maps before (a) and after (b) preprocessing.
• Figure 2: 3D deep neural network architecture for MRSI metabolic maps missing data estimation.
• Figure 3: Comparison of 2D simulated MRSI metabolic map restoration results on 20% missing data using the proposed deep learning method and traditional interpolation techniques. Beyond the differences in the ventricular regions at the center of the images, the yellow boxes highlight reconstruction discrepancies in tumor-affected areas, while the red boxes indicate notable differences in healthy brain regions. These areas illustrate the superior structural preservation and detail recovery achieved by the proposed approach.
• Figure 4: Missing data estimation by the proposed model on simulated 3D metabolic maps with 15% missing data: Sample 1 (a) and Sample 2 (b). In each sample, the coronal, sagittal, and axial views are shown from left to right, respectively, while the input, the model's prediction, and the ground truth are shown from top to bottom.
• Figure 5: Restoration results for 15% missing data estimation in 2D N-acetylaspartate (NAA) metabolic maps from real healthy subject data. The proposed model was tested on four representative samples (a–d). For each sample, the first column displays the input image with simulated missing voxels, the second column shows the corresponding reconstruction generated by the model, and the third column presents the ground truth metabolic maps for reference.