Medical imaging is widely used in current healthcare and research settings for various purposes such as diagnosis, treatment options, patient monitoring, longitudinal studies, etc. The two most commonly used imaging modalities in the United States are Computed Tomography (CT) and Magnetic Resonance Imaging (MRI). Raw images acquired via CT or MRI need to undergo a variety of processing steps prior to being used for the purposes explained above. These processing steps include quality control, noise reduction, anatomical segmentation, tissue classification, etc. However, since medical images often include millions of voxels (smallest 3D units in the image containing information) it is extremely challenging to process them manually by relying on visual inspection and the experience of trained clinicians. In light of this, the field of medical imaging is seeking ways to automate data processing. With the impressive performance of Artificial Intelligence (AI) in the field of Computer Vision, researchers in the medical imaging community have shown increasing interest in utilizing this powerful tool to automate the task of processing medical imaging data. Despite AI’s significant contributions to the medical imaging field, large cohorts of data still remain without optimized and robust AI-based tools to process images efficiently and accurately.
This thesis focuses on exploiting large cohorts of CT and MRI data to design AI-based methods for processing medical images using weakly-supervised and supervised learning strategies, as well as mathematical (and/or statistical) modeling and signal processing methods. In particular, we address four image processing problems in this thesis. Namely: 1) We propose a weakly-supervised deep learning method to automate binary quality control of diffusion MRI scans into ‘poor’ and ‘good’ quality classes; 2) We design a weakly-supervised deep learning framework to learn and detect visual patterns related to a set of different artifact categories considered in this work, in order to identify major artifact types present in dMRI volumes; 3) We develop a supervised deep learning method to classify multiple lung texture patterns with association to Emphysema disease on human lung CT scans; 4) We investigate and characterize the properties of two types of negative BOLD response elicited in human brain fMRI scans during visual stimulation using mathematical modeling and signal processing tools.
Our results demonstrate that through the use of artificial intelligence and signal processing algorithms: 1) dMRI scans can be automatically categorized into two quality groups (i.e., ‘poor’ vs ‘good’) with a high classification accuracy, enabling rapid sifting of large cohorts of dMRI scans to be utilized in research or clinical settings; 2) Type of the major artifact present in ‘poor’ quality dMRI volumes can be identified robustly and automatically with high precision enabling exclusion/correction of corrupt volumes according to the artifact type contaminating them; 3) Multiple lung texture patterns related to Emphysema disease can be automatically and robustly classified across various large cohorts of CT scans enabling investigation of the disease through longitudinal studies on multiple cohorts; 4) Negative BOLD responses of different categories can be fully characterized on fMRI data collected from visual stimulation of human brain enabling researchers to better understand the human brain functionality through studying cohorts of fMRI scans.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/92xy-a712 |
| Date | January 2022 |
| Creators | Ettehadi, Nabil |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0017 seconds