Date of Award
Doctor of Philosophy (PhD)
Tomographic image reconstruction is generally an ill-posed inverse problem. Such inverse problems are typically regularized using prior knowledge of the sought-after object property. Recently, deep neural networks have been actively investigated for regularizing image reconstruction problems by learning a prior for the object properties from training images. Deep generative models such as generative adversarial networks (GANs) have demonstrated the ability to learn object distributions comprehensively and synthesize high-quality images. This dissertation explores novel generative model-constrained reconstruction methods that employ state-of-the-art GANs in the context of ill-posed tomographic imaging problems. The symbiotic relationship between image science and deep learning to enable responsible artificial intelligence (AI) applications in medical imaging is also demonstrated. In the first part of the dissertation, an image reconstruction method is proposed (IAGAN-TV), which extends the IAGAN method introduced in ill-posed image restoration problems with an improved regularization strategy and employs a progressively growing GAN architecture in an image-adaptive framework. We demonstrate the ability of the IAGAN-TV method to recover fine structures in ill-posed image reconstruction problems, which cannot be achieved using sparsity-promoting penalties alone. The stability and generalization properties of the proposed method are established. In the second part, a formal definition of “hallucinations” is introduced in the context of image reconstruction using fundamental image science principles derived from linear operator theory. We demonstrate how the ability to define image hallucinations allows the quantification of false structures in reconstructed images, enabling preliminary assessments of deep learning-based reconstruction methods via virtual imaging trials. In the final part of the dissertation, a method to produce multiple data-consistent solutions to image reconstruction problems is proposed (PULSE++) that employs a style-based GAN architecture (StyleGAN). The PULSE++ method extends the PULSE method introduced in single-image super-resolution tasks to general ill-posed inverse problems. The proposed method improves the performance of PULSE by stabilizing the core optimization method and utilizing more accurate statistical knowledge of the StyleGAN latent space. The scalability of the PULSE++ method and its effectiveness with different practical measurement noise distributions is demonstrated. We illustrate how the ability to produce multiple data-consistent solutions using the PULSE++ method enabled new assessments of imaging systems, such as uncertainty quantification in image reconstruction.
Ulugbek S. Kamilov Mark Anastasio
Ulugbek S. Kamilov, Mark A. Anastasio, Abhinav K. Jha, Brendan Juba,
Available for download on Wednesday, May 15, 2024