Deep neural networks for mri applications

Lyu, Qing
Thumbnail Image
Other Contributors
Yan, Pingkun
Lai, Rongjie
Hahn, Juergen
Wang, Ge
Issue Date
Biomedical engineering
Terms of Use
Attribution-NonCommercial-NoDerivs 3.0 United States
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute (RPI), Troy, NY. Copyright of original work retained by author.
Full Citation
Magnetic resonance imaging (MRI) has been widely used for clinical disease diagnosis and neuroscience research since its invention. Compared with other commonly used medical imaging modalities like computed tomography and ultrasound, MRI has advantages in showing soft tissue in rich contrast, introducing no ionizing radiation during the scanning, and presenting either anatomical or physiological information through the flexible configuration of the scanning pulse sequence. It is estimated there are around 40 million MRI scans conducted in the United States every year. Despite MRI having achieved great success in recent years, it still has some limitations like long scan time and low signal-to-noise ratio. With the emergence of powerful GPU-based computing systems and the collection of large-scale open-access datasets, deep learning has achieved great progress and profoundly changed the world in multiple regions over the past decade, ranging from pattern recognition to healthcare. Numerous studies have shown the performance of neural networks superior to humans on specific tasks. However, studies on working principles of deep learning like network interpretability, generalization, and stability are still insufficient, raising safety concerns and limiting its actual deployment in real-world applications. This dissertation addresses problems in three aspects: 1) using deep learning to overcome existing MRI shortcomings; 2) using deep learning to broaden the application scope of MRI; and 3) exploring unsolved deep learning problems. For existing MRI shortcomings, we aim at 1) shortening MRI scan time without sacrificing image quality, 2) reducing cine cardiac MRI motion artifacts, and 3) correcting the error caused by T2 defocusing happened in current MRI reconstruction. Through conducting experiments on single-contrast and multi-contrast super-resolution and proposing neural networks with advanced architectures, objective functions, and optimization algorithms, we can successfully shorten MRI scan time by at least 4-fold without significant image quality degradation. In the cine cardiac MRI study, we propose a recurrent neural network with convolutions and long short-term memory. According to the feedback of collaborating clinicians, our results greatly mitigate motion artifacts and can better be used for cardiovascular diseases diagnosis. In the study trying to correct the error caused by T2 defocusing that happened in current MRI reconstruction, we propose a deep learning framework integrating the MRI data acquisition process and image reconstruction process, and optimizing both the pulse sequence and the reconstruction scheme seamlessly. According to our pilot simulation results, better MR images can be obtained. For the extension of MRI application, we propose a deep learning-based end-to-end framework that can classify brain metastases based on their primary organ sites from whole-brain MRI scans with minimal human intervention. Through designing neural networks for metastases segmentation, image modality transformation, and classification, we can classify brain metastases into five categories with high accuracy. Our study shows the potential to use a whole-brain MRI scan instead of biopsy for metastases classification in the future, which also contributes to the early diagnosis of cancer. For the exploration of unsolved deep learning problems, we conduct a study on neural network uncertainty. We propose a method to estimate neural network uncertainty so that the reliability of network results can be quantified. We then apply the proposed method to some MRI applications and show that quantifying network result uncertainty can deliver better diagnostic performance and make medical AI imaging more explainable and trustworthy.
May 2022
School of Engineering
Dept. of Biomedical Engineering
Rensselaer Polytechnic Institute, Troy, NY
Rensselaer Theses and Dissertations Online Collection
CC BY-NC-ND. Users may download and share copies with attribution in accordance with a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 license. No commercial use or derivatives are permitted without the explicit approval of the author.