MRI recovery with self-calibrated denoisers without fully-sampled data

Area of Science:

• Medical Imaging
• Machine Learning
• Image Reconstruction

Background:

• Acquiring fully sampled training data for magnetic resonance imaging (MRI) applications is often difficult.
• Existing methods may require extensive, fully sampled datasets for training image reconstruction models.

Purpose of the Study:

• To introduce ReSiDe, a novel self-supervised image reconstruction method for MRI.
• To enable image recovery solely from undersampled data, overcoming training data limitations.

Main Methods:

• ReSiDe is inspired by plug-and-play (PnP) methods but iteratively trains the denoiser during reconstruction.
• Two variations, ReSiDe-S (scan-specific) and ReSiDe-M (multi-set), were developed and compared.
• Evaluations were conducted on T1/T2-weighted brain MRI, MRXCAT phantom, and cardiac perfusion data.

Main Results:

• ReSiDe-S and ReSiDe-M demonstrated superior performance compared to other self-supervised/unsupervised methods.
• Outperformance was measured by peak signal-to-noise ratio and structural similarity index for static imaging.
• Expert scoring confirmed effectiveness in dynamic cardiac perfusion imaging.

Conclusions:

• ReSiDe is a validated self-supervised image reconstruction technique for both static and dynamic MRI.
• This method significantly benefits MRI applications with limited access to fully sampled training data.