IREM: High-Resolution Magnetic Resonance (MR) Image Reconstruction via Implicit Neural Representation

Qing Wu, Yuwei Li, Lan Xu, Ruiming Feng, Hongjiang Wei, Qing Yang, Boliang Yu, Xiaozhao Liu, Jingyi Yu, Yuyao Zhang

Abstract

High-resolution (HR) MRI is essential for clinical diagnosis but suffers from long acquisition times and motion artifacts. We present IREM (Implicit REpresentation for MRI), a novel reconstruction framework that models 3D medical images as continuous functions using coordinate-based neural networks. By training a multi-layer perceptron (MLP) with Fourier feature mapping on a patient's own low-resolution (LR) scans, IREM synthesizes high-fidelity HR images without requiring large external datasets. Our results demonstrate that IREM consistently outperforms traditional interpolation and super-resolution methods in both quantitative metrics and visual detail recovery.

Introduction

High-resolution (HR) MRI is the gold standard for medical diagnosis, yet capturing sub-millimeter 3D detail often requires scan times exceeding 30 minutes. Such long durations increase the risk of patient motion, leading to significant image artifacts. To mitigate this, clinicians frequently acquire multiple fast, "thick-slice" low-resolution (LR) scans from orthogonal planes (coronal, axial, sagittal). While these scans are faster and more robust to motion, they lack the fine structural detail necessary for complex diagnoses. Traditional super-resolution (SR) methods, such as B-spline interpolation or CNN-based deep learning, often produce blurry results or require massive, unavailable datasets for training. IREM addresses these challenges by treating image reconstruction as an implicit neural representation problem.

Methodology

IREM shifts the paradigm from discrete voxel grids to continuous function representations. The framework consists of three primary stages: spatial normalization, model optimization, and HR image reconstruction.

Overview of the IREM pipeline including spatial normalization, model optimization, and HR reconstruction.

Continuous Function Representation

Instead of a 3D array, IREM represents an image as a continuous function $I(x, y, z) = \text{Intensity}$. This is implemented using a fully-connected MLP. To overcome the "spectral bias" inherent in neural networks—where low-frequency components are learned preferentially over high-frequency details—the authors employ Fourier feature positional encoding. This maps 3D coordinates into a higher-dimensional space, enabling the network to capture sharp anatomical structures.

Self-Supervised Training

Unlike standard deep learning models, IREM is self-supervised and patient-specific. The network is trained to minimize the Mean Squared Error (MSE) between its predicted intensities and the actual intensities observed in the patient's own LR scans. This "per-subject" approach ensures the model learns the unique anatomy of the individual without needing external training data.

Experimental Results

The framework was evaluated on three datasets: simulated 7T brain scans (Dataset #A), simulated 3T scans with white matter lesions (Dataset #B), and real-world validation scans (Dataset #C).

Quantitative and Qualitative Performance

In simulations, IREM consistently achieved higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) compared to standard Super-Resolution Reconstruction (SRR) and B-spline interpolation. Qualitatively, IREM successfully recovered sharp edges and fine textures in both gray and white matter that were lost or blurred by competing methods.

Qualitative comparison of IREM against Ground Truth, SRR, and B-Spline methods across two datasets.

Signal-to-Noise Ratio (SNR) Advantage

A key insight of IREM is its ability to leverage the high SNR of thick-slice LR scans. Because each LR voxel collects more signal than a thin HR voxel, IREM effectively "harvests" this signal from multiple views. In real-world testing (Dataset #C), IREM produced HR images with better tissue contrast and lower noise than a direct HR scan, despite requiring less total scan time. The authors noted that the SNR in the LR stacks was approximately 16 times higher than the HR reference.

Discussion and Conclusion

IREM provides a robust alternative to traditional MRI acquisition by reconstructing high-quality images from fast LR scans. Its primary advantage lies in its patient-specific optimization, which avoids generalization errors common in pre-trained CNNs. However, the requirement for a per-scan "fine-training" phase introduces computational overhead that may limit immediate clinical throughput. Future work aims to optimize this training time and extend the framework to motion-prone applications such as fetal or cardiac imaging. Overall, IREM demonstrates that implicit neural representations offer a powerful, data-efficient pathway toward high-resolution medical imaging.

Summary updated

BibTeX
@inproceedings{wu2021irem,
  author = {Qing Wu and Yuwei Li and Lan Xu and Ruiming Feng and Hongjiang Wei and Qing Yang and Boliang Yu and Xiaozhao Liu and Jingyi Yu and Yuyao Zhang},
  title = {{IREM: High-Resolution Magnetic Resonance (MR) Image Reconstruction via Implicit Neural Representation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  doi = {10.1007/978-3-030-87231-1\_7},
}
Figure