Introduction: The use of deep learning methods in medical imaging enables improved image quality and potentially reduces the need for additional imaging examinations. One promising direction is the generation of synthetic CT images from PET data, which could reduce radiation exposure to patients and simplify the PET/CT examination process. Purpose: The purpose of this thesis is to investigate the feasibility of generating computed tomography images from positron maintaining diagnostic accuracy. Methods: a descriptive method was used in the thesis, with a systematic review of the literature in the PubMed, ScienceDirect, JNM and DiKUL databases. Inclusion and exclusion criteria were applied and 10 relevant articles discussing deep learning models and sCT image generation were analyzed. Results: Studies show models such as U-Net, ResNet and CycleGAN successfully generate sCT images that closely match real CT scans. Average errors were within clinically acceptable limits and the models enable a reduction in radiation dose and exposure time. Nevertheless, challenges remain in accurately reconstructing fine anatomical structures and in overall clinical validation. Discussion and conclusion: Deep learning methods have great potential for generating sCT and reducing the need for conventional CT imaging in PET/CT. Furter validation, improvement of reconstruction accuracy and standardization of procedures will be necessary for clinical application. Artificial intelligence represents a promising but not yet fully mature tool for clinical practice.
|
