Inter-slice Super-resolution of Magnetic Resonance Images by Pre-training and Self-supervised Fine-tuning

TL;DR

This paper tackles inter-slice super-resolution for MRI where isotropic resolution is desirable but HR data are scarce. It introduces a three-stage framework that combines supervised pre-training on a large video-frame interpolation task with MR-domain fine-tuning on a high-quality MR dataset and subsequent self-supervised fine-tuning on the target dataset to adapt to user-specific cases. The approach achieves superior performance compared with other self-supervised methods on knee MRI, and video pre-training accelerates learning while self-supervised fine-tuning ensures domain and patient-level adaptation, approaching supervised upper bounds. The results suggest practical applicability for improving visualization and downstream analyses in clinical MR imaging where HR data are limited.

Abstract

In clinical practice, 2D magnetic resonance (MR) sequences are widely adopted. While individual 2D slices can be stacked to form a 3D volume, the relatively large slice spacing can pose challenges for both image visualization and subsequent analysis tasks, which often require isotropic voxel spacing. To reduce slice spacing, deep-learning-based super-resolution techniques are widely investigated. However, most current solutions require a substantial number of paired high-resolution and low-resolution images for supervised training, which are typically unavailable in real-world scenarios. In this work, we propose a self-supervised super-resolution framework for inter-slice super-resolution of MR images. Our framework is first featured by pre-training on video dataset, as temporal correlation of videos is found beneficial for modeling the spatial relation among MR slices. Then, we use public high-quality MR dataset to fine-tune our pre-trained model, for enhancing awareness of our model to medical data. Finally, given a target dataset at hand, we utilize self-supervised fine-tuning to further ensure our model works well with user-specific super-resolution tasks. The proposed method demonstrates superior performance compared to other self-supervised methods and also holds the potential to benefit various downstream applications.

Figures (2)

• Figure 1: The proposed three-stage super-resolution framework, combining advantages of pre-training, fine-tuning, and self-supervised fine-tuning.
• Figure 2: Qualitative results and error maps for different self-supervised SR methods on the SKI10 dataset.