Year: 2025
Type: dissertation
Abstract: Score-based diffusion models based on stochastic processes have proven to be time-efficient solutions to denoising problems. By conditioning these stochastic processes with observed measurements using statistical inversion methods, score-based diffusion models can potentially be used to solve inversion problems effectively in terms of time. Especially in medical imaging, solving inversion problems requires long and computationally heavy imaging, for which conditional score-based diffusion models could provide more effective solutions in terms of time and computational cost. This bachelor's thesis first explores how Ornstein-Uhlenbeck stochastic processes and Bayesian inversion can be used to implement a conditional score-based diffusion model in theory. It then investigates how the conditional score-based diffusion model implemented on the basis of theory is able to solve the medical imaging associated inversion problems (Gauss blurring, Radon transformation) with certain parameter settings. The results show that the score-based diffusion model is able to solve both inversion problems with reasonably challenging parameter settings, so that the model could potentially be useful in practical applications. However, the modeling was performed with simple image data, so the performance of the model should be investigated in more detail with more complex data.
Source: LUTPub (LUT University)
Author: Anttalainen, Sami
Institution: Lappeenranta-Lahti University of Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7065482961