Deep probabilistic and generative models for x-ray based imaging and ecg
Loading...
Authors
Wiedeman, Christopher
Issue Date
2025-05
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Electrical engineering
Alternative Title
Abstract
Artificial intelligence (AI) is poised to transform modern medicine and, in particular, medical imaging, as the need for healthcare services continues to outstrip available resources worldwide. Deep learning models show promise in enhancing clinical workflows and patient outcomes through data-driven diagnoses and improved image quality. Despite the pressing need and considerable research efforts, several challenges have delayed the integration of AI into healthcare: Neural networks often lack explainability, making overconfident inferences on out-of-sample data, and are susceptible to adversarial attacks—small, targeted input perturbations capable of fooling otherwise accurate networks. Furthermore, large models are data-hungry, yet patient images and information are legally guarded by health privacy protections. Future virtual clinical trials for medical device validation also require high-quality synthetic datasets that sufficiently represent rare pathologies and population demographics. While deep generative models may address these data gaps, their outputs often lack clinical fidelity despite appearing visually convincing. Finally, many image-enhancement tasks, such as deblurring, lack the ground-truth labels necessary for supervised learning, forcing reliance on simulated degradations that can introduce artifacts when applied to real-world data. Compounding the above issues is the high-dimensional nature of most medical signals, invalidating solutions that cannot be feasibly scaled. The following dissertation investigates solutions to several of the aforementioned challenges within specific applications, primarily leveraging deep probabilistic and generative models. First, we examine how diverse deep ensembles, aided by a feature decorrelation mechanism, can improve adversarial robustness in high-dimensional tasks like electrocardiogram classification. Next, drawing inspired by the 2023 AAPM Deep Generative Modeling Challenge, we employ denoising diffusion probabilistic models to produce synthetic medical images that are realistic in both visual appearance and clinical relevance—an effort that earned first place in the competition. Finally, we adapt state-of-the-art simulation techniques to create realistic, system specific degradations to train deep deblurring models for photon-counting computed tomography. By scaling a diffusion model to 3D through a joint 2D inference process and disentangling noise from signal prior to deblurring, we successfully mitigate texture distortions and improve performance on real-world data.
Description
May2025
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY