Logarithmic Scaling of Loss Functions for Enhanced Self-Supervised Accelerated MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) offers non-invasive, high-contrast soft-tissue visualization without ionizing radiation.
• Accelerated acquisition is crucial for high-resolution MRI due to lengthy scan times.
• Self-supervised learning (SSL) methods reconstruct undersampled MRI data without requiring fully sampled ground truth.

Purpose of the Study:

• To enhance self-supervised MRI reconstruction using a novel logarithmic scaling scheme for conventional loss functions.
• To address the tendency of standard k-space domain losses to under-represent high-frequency information.
• To improve the perceptual quality and fidelity of reconstructed MRI images.

Main Methods:

• A logarithmic scaling scheme was applied to standard loss functions (e.g., L1, L2) within a self-supervised framework.
• The proposed method adaptively rescales residuals, emphasizing high-frequency components in MRI reconstruction.
• The approach is designed to be lightweight, architecture-agnostic, and easily integrated into existing pipelines.

Main Results:

• Consistent quantitative improvements were observed across public datasets when using the proposed log-scaled loss.
• The method demonstrated enhanced reconstruction fidelity compared to standard self-supervised approaches.
• Improved perceptual quality of the reconstructed MRI images was achieved.

Conclusions:

• The proposed log-scaled loss function effectively improves self-supervised MRI reconstruction.
• The method enhances both quantitative metrics and perceptual quality of accelerated MRI scans.
• This lightweight and adaptable approach offers a valuable addition to existing MRI reconstruction techniques.