[Defense] Statistical Analysis and Deep Learning Techniques for Enhancing Low Quality Magnetic Resonance Imaging
Wednesday, July 5, 2023
9:00 am - 10:00 am
In
Partial
Fulfillment
of
the
Requirements
for
the
Degree
of
Doctor
of
Philosophy
Rishabh
Sharma
will
defend
his
dissertation
Statistical
Analysis
and
Deep
Learning
Techniques
for
Enhancing
Low
Quality
Magnetic
Resonance
Imaging
Abstract
Magnetic Resonance Imaging (MRI) is a widely employed non-invasive diagnostic tool in modern medicine, offering detailed images of internal structures within the body. While MRI presents several advantages, such as safety and superior soft tissue contrast, it also possesses limitations, including reduced sensitivity and trade-offs between signal-to-noise ratio (SNR), spatial resolution, and acquisition time. To overcome these limitations, this dissertation explores the application of deep learning methods, such as UNet convolutional networks and generative adversarial networks (GANs), to enhance the quality of low-resolution and low signal-to-noise ratio MRI images. This dissertation focuses on the statistical analysis of different UNet variants and their performance in enhancing MRI image quality. The dissertation proposes the use mixed effects models to evaluate and compare the performance of UNet architectures. Mixed effects models account for image-to-image variabilities that may affect the analysis of outcomes in image analysis and is the correct method to use when repeated measurements are done of test dataset. The work also addresses the selection of appropriate loss functions in UNet training, which is crucial for accurate and robust image enhancement results. It highlights the challenge of choosing the optimal loss function based on descriptive statistics alone and emphasizes the need for human interpretation of results and consideration of image to image variations. Additionally, the impact of training data quality on UNet performance is explored, emphasizing the importance of diverse and high quality datasets for achieving reliable image enhancement. Furthermore, a patch-based learning approach is found to outperform training on entire images, producing more robust results with fewer artifacts. Utilizing patches enables a finer and localized learning process, enhancing model adaptability to different image regions. This patch-based approach is recommended as a reliable method for image enhancement. Moreover, the investigation of multicontrast MRI scenarios provides valuable insights into enhancing low quality images. Results demonstrate that patch-based learning with the DUNet architecture remains reliable even in multicontrast scenarios. Leveraging multiple contrasts in the training process improves enhancement outcomes, leading to sharper and diagnostically valuable images. Overall, this dissertation contributes to the field of medical image analysis by incorporating statistical analysis techniques, evaluating different UNet variants, exploring the role of loss functions and training data quality, and using multicontrast MRI and patch-based learning in MRI image enhancement. The findings provide insights into selecting optimal models, improving explainability, and enhancing the overall quality of MRI images, thereby contributing to better clinical outcomes and patient care.
Wednesday,
July
5,
2023
9:00AM
Online
via
MS
Teams
Dr. Nikolaos V. Tsekos, Faculty Advisor
Faculty, students and the general public are invited.
