Development of deep-learning-based denoising algorithms for fast whole-body [18F]FDG PET/CT scans

• Autor(a) principal: Silva, Luísa Castelbranco da Silveira Coelho
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/160558
• topic: [18F]FDG PET/CT; deep learning; denoising; molecular imaging; oncology; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

This study aims to assess the feasibility of reducing the acquisition time of whole-body 18Flabelled fluorodeoxyglucose ([18F]FDG) positron emission tomography coupled with computed tomography (PET/CT) scans through deep-learning-based denoising. 112 whole-body [18F]FDG PET/CT scans of patients with cancer were included. 92 were employed in the training of three convolutional neural networks: 2D, 2.5D and 3D U-Nets. Mean squared error (MSE) was the appointed loss function. The remaining 20 scans were set aside for testing. The images were acquired on a Philips Vereos Digital PET/CT scanner. From the standard-duration (70 seconds per axial field of view (AFOV)) raw data, fast scans were simulated by cropping the data to 15, 20 and 30 s/AFOV. Reconstructionwas performed on-site using the manufacturer’s protocol and following EARL1 standards. MSE, structural similarity index measure (SSIM) and intraclass correlation coefficient (ICC) were used for a voxel-wise comparison between the deep-learning-denoised (DL-denoised) fast scans and the reference images (70 s/AFOV). Signal-to-noise ratio (SNR) was computed in regions with expected uptake uniformity (liver and lungs) through the quotient between the mean standardised uptake value (SUV) and the SUV standard deviation. On a tumour basis, quantification was performed in terms of maximum SUV, mean SUV, SUV standard deviation, peak SUV, total lesion glycolysis (TLG) and metabolic tumour volume (MTV) in both the lesions in the DLdenoised and reference images. For benchmarking, Gaussian filter (GF), the state-of-the-art denoising method, was implemented and its width optimised in the training set through MSE minimisation relatively to the reference images. The voxel-wise results revealed a strong agreement between the DL-denoised 15, 20 and 30-s/AFOV-based sets and the reference images, with an ICC equal or higher than 0.985. Quantification in the liver and lungs unveiled the DL-denoised images to have higher SNR compared to the original (fast), the GF-denoised and even the reference images. Tumour quantification exposed variations in the lesions’ features that are not expected to have clinical impact, particularly in the 20 and 30-s/AFOV-based sets. Deep-learning-based denoising outperformed optimised Gaussian filter in every instance. The deep-learning-based denoising models for fast whole-body [18F]FDG PET/CT scans developed in this study proved to have potential to achieve images with clinically-suitable quantitative parameters. The 20 s/AFOV scans with post-processing with the 2.5D U-Net or the 3D U-Net seemed to be the best compromise between scan duration and image quality, compared to the 15 and 30-s/AFOV-based scans.