Date of Graduation

12-2023

Document Type

Thesis

Degree Name

Master of Science in Computer Science (MS)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

Ukash Nakarmi

Committee Member

Gauch, Susan

Second Committee Member

Zhang, Lu

Keywords

medical imaging, training data, degradation methods, model biases, deep learning

Abstract

Super-resolution has emerged as a crucial research topic in the field of Magnetic Resonance Imaging (MRI) where it plays an important role in understanding and analysis of complex, qualitative, and quantitative characteristics of tissues at high resolutions. Deep learning techniques have been successful in achieving state-of-the-art results for super-resolution. These deep learning-based methods heavily rely on a substantial amount of data. Additionally, they require a pair of low-resolution and high-resolution images for supervised training which is often unavailable. Particularly in MRI super-resolution, it is often impossible to have low-resolution and high-resolution training image pairs. To overcome this, existing methods for training super-resolution models simulate low-resolution images from the available few high-resolution images using a specific degradation method, which may not accurately represent real-world image degradation. This can lead to poor performance of the models when applied to realistic scenarios. In this work, we investigate the impact of the quality of training examples on the performance of deep learning models for super-resolution. Specifically, we hypothesize that the model is biased on the degradation nature of the training examples, and this bias can be reduced by incorporating greater variability into the training data set. To test this hypothesis, we trained 1) a simple convolution network, 2) a Dense Network, and 3) a Residual in Residual Dense Network on data sets with different types of simulated degradation, and compared their performance with models trained on a diverse data set. Our results show that the model gives better performance when the type of degradation in both the training and test sets are similar. However, the model trained on a diverse data set demonstrates greater robustness and surpasses models that were trained only on a single type of degradation. This suggests that incorporating a variety of degradation types in the training data set can enhance the performance of deep learning models for super-resolution. Furthermore, we created an application to conduct a Mean Opinion Score (MOS) Survey to get a better measure of image perceptual quality.

Download

Share

COinS