Self-Supervised Super-Resolution for Anisotropic MR Images with and Without Slice Gap.

Area of Science:

• Medical Imaging
• Image Processing
• Artificial Intelligence

Background:

• Magnetic resonance (MR) imaging often uses thick slices to reduce scan time and motion artifacts.
• Thick slices and inter-slice gaps in MR images can compromise volumetric analysis and 3D reconstruction accuracy.
• Existing super-resolution (SR) methods struggle with anisotropic MR data, especially those with slice gaps, and are prone to domain shift.

Purpose of the Study:

• To develop a self-supervised super-resolution (SR) technique capable of handling anisotropic MR images, including those with slice gaps.
• To improve the accuracy of volumetric analysis and 3D methods by enhancing MR image resolution.
• To create a robust SR method that mitigates domain shift issues inherent in data-driven approaches.

Main Methods:

• Proposed a novel self-supervised super-resolution (SR) algorithm specifically designed for anisotropic MR images.
• The method effectively addresses scenarios with both thick slices and inter-slice gaps.
• Compared the proposed SR technique against existing methods on two open-source datasets.

Main Results:

• The self-supervised SR method demonstrated significant improvements in both signal recovery and downstream task performance.
• The technique proved effective for MR images with and without slice gaps, outperforming competing methods.
• Validation across two datasets confirmed the robustness and generalizability of the proposed approach.

Conclusions:

• The developed self-supervised SR technique offers a robust solution for enhancing anisotropic MR images, particularly those with slice gaps.
• This method improves the accuracy of volumetric analysis and 3D reconstruction, crucial for clinical applications.
• The publicly available code facilitates further research and application of this advanced image processing technique.