Tipus de document

Treball de fi de grau

Data de publicació

Tots els drets reservats

Si us plau utilitzeu sempre aquest identificador per citar o enllaçar aquest document: https://hdl.handle.net/2445/231336

Deep Learning-Based Denoising of Low-Dose Monte Carlo PET Simulations

Títol de la revista

Contribució addicional

ISSN de la revista

Títol del volum

Resum

Monte Carlo simulations play a crucial role in nuclear medicine by enabling realistic modeling of PET acquisition processes. However, they are computationally expensive, and reducing acquisition time or injected tracer dose often results in increased noise and degraded image quality. This study investigates the capacity of Deep Learning architectures to address this challenge by denoising low-dose PET images. Utilizing a Monte Carlo simulation platform, we evaluated Standard and Attention-augmented 2D and 3D U-Nets across a spectrum of simulated low-dose scenarios, down to an extreme 0.01 mCi. Quantitative and qualitative assessments at the 0.255 mCi dose level demonstrated promising trends. The 3D Attention U-Net showed superior performance compared to its counterparts, successfully mitigating structural anomalies and geometric noise to yield robust anatomical coherence and a reliable metabolic quantification error under 5%. While stress-testing at 0.01 mCi exposed a strict physical boundary where the lack of structural information prevents a feasible reconstruction, these preliminary findings suggest the potential viability of attention-guided networks for safe, low-dose medical image reconstruction

Descripció

Treballs Finals de Grau de Física, Facultat de Física, Universitat de Barcelona, Curs: 2026, Tutors: Aida Niñerola Baizán, Arnau Farré Melero

Recurs relacionat

Citació

Citació

CAMPO MASCARAY, Marta. Deep Learning-Based Denoising of Low-Dose Monte Carlo PET Simulations. [consulted: 20 of November of 2026]. Available at: https://hdl.handle.net/2445/231336

Exportar metadades

JSON - METS

Compartir registre