All-in-one Deep Learning Framework for MR Image Reconstruction

Authors: Geunu Jeong, Hyeonsoo Kim, Joonyoung Yang, Kyungeun Jang, Jeewook Kim (AIRS Medical)

This research introduces a novel, all-in-one deep learning framework for MR image reconstruction, enabling a single model to enhance image quality across multiple aspects of k-space sampling and to be effective across a wide range of clinical and technical scenarios. Multi-dimensional degradation was applied to raw k-space data to generate training input. This process involved a combined application of noise addition and multiple patterns of undersampling (uniform, random, kmax, partial Fourier, elliptical), with each method being applied across a range of factors to cover extensive sampling scenarios. Contextual data, including scan parameter information, were prepared to serve as auxiliary input to address the challenges posed by the unique learning task for each training pair, which arise from the varied degradation scenarios. The expected noise reduction factor for each training pair was mathematically derived and used as additional contextual data for tunable denoising. The U-Net was modified to include an additional pathway for integrating contextual data. The proposed model enhances image quality in a multi-dimensional manner and is compatible with a broad spectrum of scenarios, including various vendors, pulse sequences, scan parameters, and anatomical regions.