Compressed Sensing Reconstruction with Zero-Shot Self-Supervised Learning for High-Resolution MRI of Human Embryos

Area of Science:

• Medical Imaging
• Developmental Biology
• Artificial Intelligence

Background:

• High-resolution magnetic resonance imaging (MRI) is crucial for studying human embryo development.
• Long scan times can limit imaging feasibility due to specimen viability and data acquisition constraints.
• Deep learning reconstruction methods offer potential for accelerating MRI acquisition.

Purpose of the Study:

• To evaluate the efficacy of zero-shot self-supervised learning (ZS-SSL) in reducing scan time for high-resolution human embryo MRI.
• To assess if ZS-SSL can maintain spatial resolution at accelerated scan rates compared to conventional methods.
• To determine the optimal acceleration factors for ZS-SSL in this application.

Main Methods:

• Simulations using numerical phantoms to assess spatial resolution at various acceleration factors (AF) and signal-to-noise ratios (SNR).
• Quantification of resolution using the Sparrow criterion and comparison of ZS-SSL with compressed sensing (CS).
• Experimental high-resolution MRI (30 μm)³ of a human embryo (Carnegie stage 21) using retrospective and prospective undersampling at AF = 4 and 8.

Main Results:

• ZS-SSL demonstrated superior spatial resolution preservation compared to CS, especially at lower SNRs.
• At AF = 4, ZS-SSL achieved image quality comparable to fully sampled data.
• Experimental imaging at AF = 4 allowed clear visualization of embryonic structures; AF = 8 resulted in reduced clarity.

Conclusions:

• ZS-SSL enables substantial scan time reduction in high-resolution human embryo MRI while preserving spatial resolution at AF = 4 (SNR > 15).
• This acceleration-image quality trade-off is advantageous for time-sensitive studies and limited specimen availability.
• The method facilitates efficient ultra-high-resolution data acquisition, supporting the creation of detailed developmental atlases.