A two-stage PET/MRI attenuation-correction workflow for pseudo-CT and AC PET synthesis with GUI-based inference

Quantitative positron emission tomography/magnetic resonance imaging (PET/MRI) requires reliable attenuation correction, but MRI does not directly encode attenuation-related tissue information. This study implemented and internally evaluated a graphical user interface (GUI)-supported two-stage brain PET/MRI workflow for pseudo-computed tomography (pseudo-CT) generation and PET attenuation-correction synthesis using the CERMEP-IDB-MRXFDG cohort. The objective was not to introduce a new neural-network architecture, but to evaluate an attenuation-informed workflow that sequentially combines MRI-based pseudo-CT generation, attenuation-map estimation, and NAC PET correction. Thirty-seven normal adult subjects were analyzed using subject-level five-fold cross-validation. A first-stage two-dimensional U-Net predicted pseudo-CT from T1-weighted MRI. The pseudo-CT output was converted into an approximate 511-keV linear attenuation coefficient map (μ-map) and combined with non-attenuation-corrected (NAC) PET as input to a second U-Net for predicted attenuation-corrected (AC) PET synthesis. Patient-level pseudo-CT errors, image-similarity metrics, Bland-Altman agreement, and exploratory standardized uptake value (SUV) metrics were evaluated. The pseudo-CT model achieved whole-image mean absolute error (MAE) of 72.26 ± 20.11 HU and structural similarity index measure (SSIM) of 0.881 ± 0.024, with bone-dominated residual error. Predicted AC PET achieved 76.02% lower normalized MAE and 73.33% lower normalized root mean square error (RMSE) than NAC PET relative to reference AC PET. Exploratory SUV analysis showed mean percentage errors of 8.66% for SUVmean and 22.08% for SUVmax, with greater inter-subject variability in SUVmax. These findings support internal CERMEP feasibility, while nested and external validation remain necessary.