Date of Award
5-14-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Deep generative models (DGMs) have tremendous potential for several biomedical imaging applications such as data augmentation, image reconstruction, and image denoising. However, the deployment of DGMs in real-world biomedical imaging workflows without domain-relevant evaluations can jeopardize patient health and well-being. The evaluation of DGMs in biomedical imaging is challenging due to several factors: requirement of domain expertise for visual inspection, lack of a mathematically defined ground truth, and the unclear relevance of popular evaluation measures adopted from the computer vision literature. Given these challenges, one way to evaluate DGMs is via purposefully designed synthetic data. In this thesis, two frameworks for the evaluation of DGMs are proposed based on the idea of assessing reproducible ``spatial context”. Context is defined as domain-specific external knowledge that manifests as conditional co-occurrences of specific pixel arrangements in an image. In the first of two frameworks, stochastic context models were purposefully designed to encode and assess the reproducibility of explicitly prescribed spatial context. Context was encoded in these models as contextual attributes such as per-image feature prevalence, feature-specific intensity distribution, and prescribed texture. In the second evaluation framework, a more complex dataset: a stochastic model of the human female breast was adapted to evaluate DGMs for reproducible spatial context that arises implicitly due to structural variations in anatomy. All designed datasets are made publicly available to aid the benchmarking of novel and emerging DGMs. The designed evaluation frameworks were employed to assess diffusion models, which are state-of-the-art DGMs and have been claimed to substantially outperform the other major DGM family: generative adversarial networks, in terms of visual image quality and popular evaluation measures. It was found that diffusion models hold promise for data augmentation tasks, but errors may occur in the generation of multiple contextual attributes, and that popular evaluation measures do not capture these contextual errors. From all studies, it was found that no modern DGM perfectly reproduced the expected spatial context. This highlights the need for further development of domain-specific DGMs as well as domain-relevant evaluation methods to ensure the safe and beneficial translation of DGM-based methods to real-world workflows in biomedical imaging. (Copyright statement: A part of this work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.)
Language
English (en)
Chair
Baranidharan Raman
Committee Members
Mark Anastasio