Highly accelerated multishot echo planar imaging through synergistic machine learning and joint reconstruction.

Area of Science:

• Magnetic Resonance Imaging
• Medical Imaging
• Computational Imaging

Background:

• Single-shot echo planar imaging (EPI) suffers from distortion artifacts, limiting high-resolution applications.
• Multishot EPI (msEPI) can reduce artifacts but is challenged by shot-to-shot phase variations.
• Existing methods struggle to achieve high-quality msEPI due to phase mismatches.

Purpose of the Study:

• To develop a combined machine learning (ML) and physics-based framework for accelerated, navigator-free msEPI.
• To demonstrate the framework's utility in high-resolution structural and diffusion imaging.

Main Methods:

• A deep learning model was used to generate an initial artifact-reduced image.
• Image phase variations were estimated from the ML-derived image.
• Joint virtual coil sensitivity encoding (JVC-SENSE) reconstruction incorporated estimated phase variations for improved image reconstruction.

Main Results:

• Achieved R_inplane × multiband (MB) = 8 × 2-fold acceleration with 2 EPI shots for multiecho imaging, enabling whole-brain T2 and T2* mapping at 1x1x3 mm³ resolution in 8.3 seconds.
• Enabled high-resolution diffusion imaging with high geometrical fidelity using 5 shots at R_inplane × MB = 9 × 2-fold acceleration.
• Extended the MUSSELS reconstruction technique for simultaneous multislice encoding as input to the ML network.

Conclusions:

• The combined ML and JVC-SENSE approach enables navigator-free msEPI at unprecedented acceleration factors with fewer shots.
• This framework demonstrates reduced vulnerability to generalizability issues and better acceptance compared to end-to-end ML methods.