Magnetic resonance image (MRI) is a widely used non-invasive radiation-free imaging technique that uniquely provides structural and functional information for disease detection, diagnosis, and treatment planning. However, the conventional MRI imaging techniques are typically slow and low in spatial or time resolution, resulting in long scan times and more susceptibility to motion artifacts. Moreover, a fast MRI scan usually comes in a low spatial resolution, making it less desirable for clinical application. A recently proposed technique, Multi-tasking MRI (MTMRI), significantly improves the scan efficiency with high temporal resolution. Nevertheless, the iterative reconstruction requires a lot of computational resources and takes a long time to process, making it challenging to fit in the clinical routine. Additionally, when doing image post-processing with MRI, despite MRI providing a good contrast of soft tissues, the variety in weighted contrast MRI’s intensity values makes it challenging to extract image features compared with other quantitative imaging techniques.
The most significant contribution of this dissertation's work is to address the three limitations above by developing a unified multi-purpose structure with deep-learning (DL) techniques. We achieved three primary goals in three different areas: 1) A general framework for highly accelerated MRI scanning without sacrificing spatial resolution, 2) reduce reconstruction time for motion-resolved free-breathing MRI technique, 3) accurately fully automated segmentation for abdominal MRI for fast image post-processing. All technical improvements utilize DL techniques to improve MRI in different aspects: to improve image quality in fast MRI scans, reduce reconstruction time in motion-resolved MRI, and reduce tedious human labors in abdominal MRI.
First, a DL-based Super-Resolution (SR) technique is developed and evaluated in both brain MRI and coronary MR Angiography (MRA). SR can recover the image quality and structural details from a 4x and 16x low-resolution fast MRI scan. For brain MRI, several SR networks have been developed. The proposed network (mDCSRN) has successfully recovered the brain structural details from a 4x low-resolution fast scan. It is developed and evaluated on an open access high-resolution T1w brain MRI with 1131 healthy volunteers. Quantitative results show that it can achieve 4x acceleration in scan while keeping similar image quality. For coronary MRA, introducing a domain adaptive network (DRAGAN) jointly trained on both coronary and brain MRA to overcome catastrophic failures commonly in training a GAN in a small dataset, we successfully accelerated the MRA acquisition by a factor of 16.
Second, DL networks are developed to accelerate the reconstruction of a 5-dimensional (5D) Multitasking MRI (MTMRI). The MTMRI is a respiratory and cardiac-motion-resolved, high-temporal-resolution technique that provides quantitative T1 mapping. However, the massive size of many dynamic MRI problems prevents deep learning networks from directly exploiting global temporal relationships. By applying deep neural networks inside a priori calculated temporal feature spaces, we enable deep learning reconstruction with global temporal modeling even for image sequences with >40,000 frames. One proposed variation of our approach using dilated multi-level Densely Connected Network (mDCN) speeds up feature space coordinate calculation by 3000x compared to conventional iterative methods, from 20 minutes to 0.39 seconds. Thus, the combination of low-rank tensor and deep learning models makes large-scale dynamic MRI feasible and practical for routine clinical application.
Third, we developed Automated deep Learning-based Abdominal Multi-Organ segmentation (ALAMO) technique based on 2D U-net and a densely connected network structure with tailored design in data augmentation and training procedures. The model takes in multi-slice MR images and generates the output of segmentation results. 3.0-Tesla T1 VIBE (Volumetric Interpolated Breath-hold Examination) images of 102 subjects were used in our study. Ten OARs were studied, including the liver, spleen, pancreas, left/right kidneys, stomach, duodenum, small intestine, spinal cord, and vertebral bodies. ALAMO generated segmentation labels in good agreement with the manual results. Specifically, among the 10 OARs, 9 achieved high Dice Similarity Coefficients (DSCs) in the range of 0.87-0.96, except for the duodenum with a DSC of 0.80. Overall, the ALAMO model matched the state-of-the-art techniques in performance.
