Deep learning approaches to magnetic resonance imaging

Abstract

As a powerful and effective medical imaging technology, magnetic resonance imaging (MRI) has gained widespread applications in both clinical diagnosis and preclinical research. In the past few years, emerging deep learning techniques offer the unprecedented opportunity to revolutionize the whole MRI processing chain, from data acquisition to image reconstruction, from automatic segmentation to diagnosis. The main research scope of this dissertation is to advance deep learning approaches for MRI reconstruction and diagnosis. Firstly, a complex-valued deep learning method was developed for general partial Fourier (PF) reconstruction. It was an unrolled convolutional neural network (CNN) that iteratively reconstructed PF sampled data and enforced data consistency. By utilizing both magnitude and phase characteristics in large-scale complex image datasets, the proposed method could effectively reconstruct MR data even at low PF fractions, yielding high-fidelity magnitude and phase images and outperforming iterative POCS methods without noticeable noise amplification even for highly PF reconstruction. The proposed method presents a valuable alternative to conventional PF reconstruction, especially for phase-sensitive 2D or 3D MRI applications. Moreover, it can be further extended to 2D PF reconstruction. Secondly, deep learning for joint multi-slice MR image reconstruction from incomplete k-space with complementary undersampling across adjacent slices was developed, including two regularly implemented PF undersampling and uniform undersampling. To exploit the structural and phase similarity across adjacent slices, a slice fusion block was introduced with learnable parameters to investigate the correlation between the target slice and its corresponding adjacent slices for joint multi-slice image reconstruction. Besides, a slice shift block was further proposed to suppress aliasing artifacts in regular uniform undersampling, where the periodic aliasing was only dependent on the acceleration factor. By utilizing the structural and phase similarity in adjacent slices together with a complementary sampling strategy, the proposed method enabled accurate MR image reconstruction in regular PF and uniform undersampling with few or even no calibration data, which can improve acquisition efficiency and flexibility in practice. Thirdly, a new MR diagnostic paradigm was presented, where the pathology detection and characterization can be performed directly from extremely sparse MR k-space data with high sensitivity and specificity. Ideally, rapid screening was desired in the first place to determine whether any pathology exists by sampling a few k-space lines. However, reconstructing high-fidelity images from such extremely sparse k-space data or noisy MR data inevitably suffered from extremely severe artifacts, which can cause detection bias and affect the adoption of rapid screening. In contrast, the input of the proposed sparse k-space detection required a much smaller data dimension, which can lead to a less complex but more robust model. Moreover, it yielded only slightly degraded performance without overfitting compared to the fully-sampled k-space pathology detection and characterization. The concept of rapid pathology detection in k-space presents an alternative MR diagnostic paradigm that may lead to potentially more powerful new tools for automatic and effective pathology screening, detection, and characterization. In conclusion, the proposed approaches through deep learning make MRI reconstructions and diagnoses more efficient, flexible, and simple in practice.