Patch-based Diffusion Models for PET Sinogram denoising

Sammanfattning

Positron Emission Tomography (PET) is a widely used imaging technique for measuring metabolic activity, but the raw data it produces is often degraded by Poisson noise due to the nature of the decay and measurement process. This thesis investigates the use of score-based diffusion models for denoising sinograms. A time-conditioned U-Net is trained to estimate the score function, which is the gradient of the log-probability of the data and it is used to solve the reverse-time stochastic differential equation that defines the diffusion process. To reduce computational demands, we explore patch-based approaches, including square, angle-wise, and sinusoidal patches, with overlapping and regularization strategies to improve consistency across patches. Experiments are conducted on synthetically generated phantoms projected using a modified minPET-3 geometry via the parallelproj package. Quantitative evaluation using pSNR, SSIM, and MSE shows that patch-based score models significantly improve denoising performance over the noisy baseline. The square patching model, angle-wise with and without regularization, perform similarly well, but none match the full-sized sinogram model.