Adaptive Local Neighborhood-based Neural Networks for MR Image Reconstruction from Undersampled Data

TL;DR

This work addresses MR image reconstruction from undersampled k-space by introducing LONDN-MRI, which locally learns reconstruction networks on adaptively selected neighboring training images during inference. The method uses a MoDL-based unrolled reconstruction and alternates between neighbor search and local training, effectively framing the process as a bilevel optimization where the neighborhood C guides the network parameters θ. Across multiple datasets and undersampling factors, LONDN-MRI yields higher image quality than globally trained models and other scan-adaptive methods, closely approaching oracle performance in many cases. The approach offers robust, scan-specific adaptation with modest training data and computational overhead, improving generalizability to varying masks, contrasts, and pathologies with potential clinical impact.

Abstract

Recent medical image reconstruction techniques focus on generating high-quality medical images suitable for clinical use at the lowest possible cost and with the fewest possible adverse effects on patients. Recent works have shown significant promise for reconstructing MR images from sparsely sampled k-space data using deep learning. In this work, we propose a technique that rapidly estimates deep neural networks directly at reconstruction time by fitting them on small adaptively estimated neighborhoods of a training set. In brief, our algorithm alternates between searching for neighbors in a data set that are similar to the test reconstruction, and training a local network on these neighbors followed by updating the test reconstruction. Because our reconstruction model is learned on a dataset that is in some sense similar to the image being reconstructed rather than being fit on a large, diverse training set, it is more adaptive to new scans. It can also handle changes in training sets and flexible scan settings, while being relatively fast. Our approach, dubbed LONDN-MRI, was validated on multiple data sets using deep unrolled reconstruction networks. Reconstructions were performed at four fold and eight fold undersampling of k-space with 1D variable-density random phase-encode undersampling masks. Our results demonstrate that our proposed locally-trained method produces higher-quality reconstructions compared to models trained globally on larger datasets as well as other scan-adaptive methods.

Figures (20)

• Figure 1: Flowchart of the proposed LONDN-MRI scheme with a specific unrolled reconstruction network. The denoising network could be for example a U-Net or the recent DIDN.
• Figure 2: Comparison of MoDL with UNet denoiser trained globally vs. using the proposed LONDN-MRI scheme (1 iteration). Reconstruction metrics are shown across training set sizes at 4x and 8x undersampling.
• Figure 3: Undersampling masks used in our experiments: (a) fourfold undersampled 1D Cartesian phase-encoded; and (b) eightfold undersampled 1D Cartesian phase-encoded. The masks were zero-padded for slightly larger images.
• Figure 4: Comparison of image reconstructions with different methods at 8x undersampling. The global and LONDN-MRI methods use the MoDL architecture with UNet denoiser with 1000 training images. The inset panel on the top left in each image corresponds to a section of interest in the image (shown by the red bounding box), while the inset panel on the top right corresponds to the error map with respect to the ground truth.
• Figure 5: Same comparisons/setup as Fig. \ref{['fig:denoised_imgs_zoomed1']}, but at 4x undersampling. The supervised methods used MoDL architecture with UNet denoiser ($3000$ training images).