Accelerating Magnetic Resonance Imaging via Deep Neural Networks in full-data and limited-data regimes

  • 13



Name of the Speaker: Mr. Amrit Kumar Jethi (EE18S037)
Guide: Dr. Mohanasankar Sivaprakasam
Date/Time: January 13th 2023 (Friday), 2:00 PM

Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool for a variety of disorders, but its utility is often limited by its slower speed and higher cost compared to modalities like CT or X-Rays. Reducing the time required for a scan would decrease the cost of MR imaging and allow for acquiring scans in situations where a patient cannot stay still for the scan duration. To accelerate the imaging time, MR scan data is sub-sampled, which brings aliasing artifacts in the image domain. Recently, artificial neural networks (ANN) have received attention in medical imaging particularly in producing artifact free images from sub-sampled scans. However, deep learning methods have certain limitations.

To begin with, the existing networks haven’t simultaneously utilized the information available in image and frequency domain. Second, though a large number of models have been proposed for multi-channel MRI reconstruction, these models need to be compared on a common platform and studied for generalizability when these models are tested on datasets acquired with different coils. Third, fully sampled data needed for supervised training of models is difficult to acquire. Algorithms need to be developed to learn useful features from easily available under-sampled data. Finally, there is a need to verify how under-sampled data can be used to enhance the generalizability of deep learning models. In this work, we develop networks to address the limitations mentioned earlier.

First, we propose Dual-Encoder-Unet which takes in both zero filled k-space and under sampled image as input and simultaneously optimizes both the domain data for reconstruction. Second, we participated in a global Multi-channel MR Image Reconstruction challenge and proposed a deep learning based cascaded architecture and evaluated its performance across different models. Third, we propose a self-supervision-based pretext learning algorithm, inspired from autoencoder for efficient feature extraction from under-sampled data. Finally, we propose three different pretext learning algorithms for multi-channel MR Image reconstruction and evaluated how pre-training can enhance the generalizability of the models. We have extensively validated all the proposed methods with suitable experiments.