MR Image Reconstruction using Deep Learning Techniques
Abstract
Magnetic Resonance Imaging (MRI) is a powerful medical diagnostic technique that provides images with a high Signal-to-Noise Ratio (SNR) and contrast information. One key parameter in clinical MRI is data acquisition time. MRI needs to be accelerated in terms of data acquisition speed and image reconstruction time. This led to the development of parallel MRI, where an array of multiple channel receiver coils is used to acquire the under-sampled k-space data. Later, advance image reconstruction algorithms are used to get the artifact free solution images. MRI data acquisition has been further accelerated in the recent past using Compressed Sensing MRI (CS-MRI) which can reconstruct images from highly under-sampled k-space data at the cost of long image reconstruction time.
Recently, researchers have used deep learning algorithms to accelerate MRI data acquisition process. The deep learning algorithms outperform the conventional MR image reconstruction algorithms in terms of computational complexity and image reconstruction time. However, the performance of end-to-end deep learning algorithms is limited by the lack of training dataset and generalization capabilities of the trained neural network.
In the first part of this thesis, a method has been proposed to cope with the data scarcity and generalization issues of the trained neural network in MR image reconstruction. Firstly, the influence of magnetic field strength of a scanner, anatomy and acceleration factors (AFs) on the generalization of a trained neural network has been assessed for MR image reconstruction problem. Later, a transfer learning approach has been proposed to reconstruct an image acquired from a different target domain dataset. It has been shown that a neural network initially trained on the human brain images obtained from 1.5T scanner can be used to reconstruct human brain images obtained from 3T scanner, cardiac images and the images under-sampled by different AFs, respectively, via end-to-end fine tuning with less training time.
In the second stage, this thesis provides an insight into the generalization issues of the trained neural network for deep learning-based receiver coil sensitivity maps. Firstly, generalization capabilities of a pre-trained neural network are thoroughly assessed for the sensitivity map estimation of receiver coils of a scanner with a different magnetic field strength; considering the same and different number of receiver coils in the source and target domain datasets. Later, transfer learning is proposed to accurately estimate the sensitivity maps of the receiver coils of different magnetic field strength scanners. It has been shown that a neural network initially trained on 8-channel receiver coil sensitivity maps of 1.5T scanner can be used to estimate the 8 or 12-channel receiver coil sensitivity maps of 3T scanner via end-to-end fine tuning. The receiver coil sensitivity maps estimated by the proposed method provide successful reconstruction in SENSitivity Encoding (SENSE).
At the third stage, a novel channel compression method has been proposed to deal with the issues that arise due to increase in number of independent channels in the receiver coils array. The handling and processing of massive MRI data from a large number of multi-channel receiver coils array limits the performance of conventional reconstruction algorithms in terms of image reconstruction time, computational complexity and memory requirements. Deep learning-based channel compression methods have been proposed in the image domain and k-space domain, separately, before CS-MRI reconstruction; without undermining the benefits of a large number of physical channels. The reconstructed images show that the proposed methods allow channel compression with minimum compression losses, retain the spatial information of virtual coils and provide high quality reconstructed images.
In this thesis, different deep learning algorithms have been proposed to provide an overall speedup in the reconstruction process of MR images without compromising their quality. Moreover, the shortcomings of end-to-end deep learning algorithms e.g. lack of training dataset for MR image reconstruction problems have also been tackled with the help of transfer learning. It is anticipated that the proposed deep learning algorithms would not only replace the computationally expensive and time-consuming conventional reconstruction algorithms but would also provide practical gains in the clinical applications of MRI.
Post a Comment