LONDN-MRI: Adaptive Local Neighborhood-Based Networks for MR Image Reconstruction from Undersampled Data

Abstract

There has been much interest in machine learning based methods for MR image reconstruction from undersampled k-space data. This paper presents a method for MR image reconstruction based on rapidly fitting neural networks to adaptively estimated neighbor-hoods within a larger training set. The weights of the network are learned only on training examples that are specified in the ‘vicinity’ or local neighborhood of a test reconstruction. The algorithm (dubbed LONDN-MRI) alternates between estimating the neighbors of the reconstructed test image and performing (local) network training and updating the test reconstruction. Rather than attempting to fit a model once to the entire dataset, our proposed method allows for learning models that are more tailored to the input test data, and therefore more flexible to the choice of undersampling patterns or anatomy. It also easily accommodates modifications to training sets. We used the recent MoDL (deep unrolled) network and the FastMRI dataset for testing our approach. We present reconstruction results for fourfold and eightfold undersampling of multi-coil data using 1D variable-density random phase-encode sampling masks. When trained locally, our method yields reconstructions of better quality compared to models learned globally on larger datasets.