2024 Theses Doctoral
Novel Image Acquisition and Reconstruction Methods: Towards Autonomous MRI
Magnetic Resonance Imaging (MR Imaging, or MRI) offers superior soft-tissue contrast compared to other medical imaging modalities. However, access to MRI across developing countries ranges from prohibitive to scarcely available. The lack of educational facilities and the excessive costs involved in imparting technical training have resulted in a lack of skilled human resources required to operate MRI systems in developing countries.
While diagnostic medical imaging improves the utilization of facility-based rural health services and impacts management decisions, MRI requires technical expertise to set up the patient, acquire, visualize, and interpret data. The availability of such local expertise in underserved geographies is challenging. Inefficient workflows and usage of MRI result in challenges related to financial and temporal access in countries with higher scanner densities than the global average of 5.3 per million people.
MRI is routinely employed for neuroimaging and, in particular, for dementia screening. Dementia affected 50 million people worldwide in 2018, with an estimated economic impact of US $1 trillion a year, and Alzheimer’s Disease (AD) accounts for up to 60–80% of dementia cases. However, AD-imaging using MRI is time-consuming, and protocol optimization to accelerate MR Imaging requires local expertise since each pulse sequence involves multiple configurable parameters that need optimization for acquisition time, image contrast, and image quality. The lack of this expertise contributes to the highly inefficient utilization of MRI services, diminishing their clinical value.
Augmenting human capabilities can tackle these challenges and standardize the practice. Autonomous and time-efficient acquisition, reconstruction, and visualization schemes to maximize MRI hardware usage and solutions that reduce reliance on human operation of MRI systems could alleviate some of the challenges associated with the requirement/absence of skilled human resources.
We first present a preliminary demonstration of AMRI that simplifies the end-to-end MRI workflow of registering the subject, setting up and invoking an imaging session, acquiring and reconstructing the data, and visualizing the images. Our initial implementation of AMRI separates the required intelligence and user interaction from the acquisition hardware. AMRI performs intelligent protocolling and intelligent slice planning. Intelligent protocolling optimizes contrast value while satisfying signal-to-noise ratio and acquisition time constraints. We acquired data from four healthy volunteers across three experiments that differed in acquisition time constraints. AMRI achieved comparable image quality across all experiments despite optimizing for acquisition duration, therefore indirectly optimizing for MR Value – a metric to quantify the value of MRI. We believe we have demonstrated the first Autonomous MRI of the brain. We also present preliminary results from a deep learning (DL) tool for generating first-read text-based radiological reports directly from input brain images. It can potentially alleviate the burden on radiologists who experience the seventh-highest levels of burnout among all physicians, according to a 2015 survey.
Next, we accelerate the routine brain imaging protocol employed at the Columbia University Irving Medical Center and leverage DL methods to boost image quality via image-denoising. Since MR physics dictates that the volume of the object being imaged influences the amount of signal received, we also demonstrate subject-specific image-denoising. The accelerated protocol resulted in a factor of 1.94 gain in imaging throughput, translating to a 72.51% increase in MR Value. We also demonstrate that this accelerated protocol can potentially be employed for AD imaging.
Finally, we present ArtifactID – a DL tool to identify Gibbs ringing in low-field (0.36 T) and high-field (1.5 T and 3.0 T) brain MRI. We train separate binary classification models for low-field and high-field data, and visual explanations are generated via the Grad-CAM explainable AI method to help develop trust in the models’ predictions. We also demonstrate detecting motion using an accelerometer in a low-field MRI scanner since low-field MRI is prone to artifacts.
In conclusion, our novel contributions in this work include: i) a software framework to demonstrate an initial implementation of autonomous brain imaging; ii) an end-to-end framework that leverages intelligent protocolling and DL-based image-denoising that can potentially be employed for accelerated AD imaging; and iii) a DL-based tool for automated identification of Gibbs ringing artifacts that may interfere with diagnosis at the time of radiological reading.
We envision AMRI augmenting human expertise to alleviate the challenges associated with the scarcity of skilled human resources and contributing to globally accessible MRI.
Geographic Areas
Subjects
Files
- Ravi_columbia_0054D_18456.pdf application/pdf 6.88 MB Download File
More About This Work
- Academic Units
- Biomedical Engineering
- Thesis Advisors
- Vaughan, John Thomas Jr.
- Geethanath, Sairam
- Degree
- Ph.D., Columbia University
- Published Here
- July 3, 2024