Super Resolution of Computed Tomography Images using 3D Machine Learning Methods

Abstract

Super resolution plays a significant role in a number of different domains, serving to
enhance image quality. There is considerable potential for super resolution to be applied
in the medical and industrial fields, particularly in facilitating diagnoses for doctors and
streamlining the inspection of industrial components. The application of super resolution
is particularly beneficial in reducing the time required for high resolution X-ray scans of
human subjects. Furthermore, the high resolution scanning of manufactured industrial
components is constrained by the physical dimensions of the specimen, which can be
circumvented through super resolution reconstruction of low resolution scans.
In this work, a comprehensive review of the current literature on super resolution
is conducted and the fundamental principles of machine learning are summarized. It
is established that deep learning and in particular U-Net network architectures, along
with generative adversarial networks (GANs), represent the current state-of-the-art in
super resolution techniques. A more detailed examination of the former two subjects is
provided in this thesis.
In this thesis, micro-computed tomography (CT) images of bone samples and other
industrial specimens are used at both low and high resolutions. 3D deep learning super resolution methods are implemented to learn the mapping between the low and
high resolution samples. In order to identify potential hyperparameters for a generalized model, a number of different model parameters are evaluated using the learned
perceptual image patch similarity (LPIPS), structural similarity index measure (SSIM)
and peak signal-to-noise ratio (PSNR). Additionally, cross validation is applied in order
to obtain a more robust estimation of the empirically best model. The best evaluated
model configuration incorporates a combined loss comprising a 0.9-weighted LPIPS and
a 0.1-weighted SSIM loss function. The artificial neural network (ANN) architecture,
which scored the best results in the experiments, is identified to be an adapted variant
of the BasicUNet network, which is implemented in the medical open network for artificial intelligence (MONAI) library. The optimal model configuration reaches a LPIPS
score of 0.1646 (baseline: 0.7729), a SSIM score of 0.8082 (baseline: 0.7473) and a PSNR
of 22.9385 (baseline: 20.2103). Further experiments with other datasets indicate that
the proposed approach can be applied to data with bone-like structures and even to
completely unrelated CT content.

Date of Award	2024
Original language	English (American)
Supervisor	Ulrich Bodenhofer (Supervisor)