Date of Award
Doctor of Philosophy (PhD)
It has been advocated to use objective measures of image quality (IQ) for assessing and optimizing medical imaging systems. Objective measures of IQ quantify the performance of an observer at a specific diagnostic task. Binary signal detection tasks and joint signal detection and localization (detection-localization) tasks are commonly considered in medical imaging. When optimizing imaging systems for binary signal detection tasks, the performance of the Bayesian Ideal Observer (IO) has been advocated for use as a figure-of-merit (FOM). The IO maximizes the observer performance that is summarized by the receiver operating characteristic (ROC) curve. When signal detection-localization tasks are considered, the IO that implements a modified generalized likelihood ratio test (MGLRT) maximizes the observer performance as measured by the localization ROC (LROC) curve. However, computation of the IO test statistic generally is analytically intractable. To address this difficulty, sampling-based methods that employ Markov-Chain Monte Carlo (MCMC) techniques have been proposed. However, current applications of MCMC methods have been limited to relatively simple stochastic object models (SOMs). When the IO is difficult or intractable to compute, the optimal linear observer, known as the Hotelling Observer (HO), can be employed to evaluate objective measures of IQ. Although computation of the HO is easier than that of the IO, it can still be challenging or even intractable because a potentially large covariance matrix needs to be estimated and subsequently inverted. In the first part of the dissertation, we introduce supervised learning-based methods for approximating the IO and the HO for binary signal detection tasks. The use of convolutional neural networks (CNNs) to approximate the IO and the use of single layer neural networks (SLNNs) to directly estimate the Hotelling template without computing and inverting covariance matrices are demonstrated. In the second part, a supervised learning method that employs CNNs to approximate the IO for signal detection-localization tasks is presented. This method represents a deep-learning-based implementation of a MGLRT that defines the IO decision strategy for signal detection-localization tasks. When evaluating observer performance for assessing and optimizing imaging systems by use of objective measures of IQ, all sources of variability in the measured image data should be accounted for. One important source of variability that can significantly affect observer performance is the variation in the ensemble of objects to-be-imaged. To describe this variability, a SOM can be established. A SOM is a generative model that can produce an ensemble of simulated objects with prescribed statistical properties. In order to establish a realistic SOM, it is desirable to use experimental data. Generative adversarial networks (GANs) hold great potential for establishing SOMs. However, images produced by imaging systems are affected by the measurement noise and a potential reconstruction process. Therefore, GANs that are trained by use of these images cannot represent SOMs because they are not established to learn object variability alone. An augmented GAN architecture named AmbientGAN that includes a measurement operator was proposed to address this issue. However, AmbientGANs cannot be immediately implemented with advanced GAN training strategies such as progressive growing of GANs (ProGANs). Therefore, the ability of AmbientGANs to establish realistic and sophisticated SOMs is limited. In the third part of this dissertation, we propose a novel deep learning method named progressively growing AmbientGANs (ProAmGANs) that incorporates the advanced progressive growing training procedure and therefore enables the AmbientGAN to be applied to realistically sized medical image data. Stylized numerical studies involving a variety of object ensembles with common medical imaging modalities are presented. Finally, a novel sampling-based method named MCMC-GAN is developed to approximate the IO. This method applies MCMC algorithms to SOMs that are established by use of GAN techniques. Because the implementation of GANs is general and not limited to specific images, our proposed method can be implemented with sophisticated object models and therefore extends the domain of applicability of the MCMC techniques. Numerical studies involving clinical brain positron emission tomography (PET) images and brain magnetic resonance (MR) images are presented.
Joseph A. O'Sullivan, Mark Anastasio
Ayan Chakrabarti, Abhinav K. Jha, Ulugbek S. Kamilov, Hua Li