부산대학교 도서관

Currently, direct methods for PET attenuation correction are typically based on NAC(Non Attenuation Correction) PET directly generating AC(Attenuation Correction) PET. However, these direct methods lack the tissue structural information from CT. To introduce CT information into the process of NAC PET directly generating AC PET, we propose a CT-guided deep learning framework. The aim is to learn the mapping relationship between attenuation-corrected NAC PET and non-attenuation-corrected AC PET images to generate PET attenuation maps. Additionally, we generate pseudo-CT images for medical observation. First, a PET-generated CT pre-training model is introduced. During training, based on sAC PET(synthetic Attenuation Correction PET), CT is generated, guided by a discriminator to aid the generator. Subsequently, quantitative metrics evaluating the training process based on sCT(synthetic CT) further assess sAC PET, avoiding the interference of noise information in sAC PET during evaluation. Finally, quantitative experiments on test data and visualization results demonstrate that the proposed CT-guided generated AC PET exhibits more similar morphological structures compared to AC PET generated without CT guidance.