Self-Supervised Learning for Full-Count PET Image Recovery with Anatomical MRI Guidance Using 18 F-FDG and 18 F-VAT PET/MRI

Positron Emission Tomography (PET) is a powerful molecular imaging modality. Low-count PET can reduce radiation dose and/or scan time, but conventional reconstruction methods often yield insufficient image quality for clinical use. Supervised deep learning methods can improve image quality but require paired low- and full-count datasets. Recent self-supervised approaches, including blind-spot networks like Noise2Void (N2V), avoid this need but depend heavily on loss function design. In this study, we propose a self-supervised PET denoising framework guided by anatomical Magnetic Resonance Imaging (MRI). A three-dimensional (3-D) dilated convolutional network was adapted to take T1-weighted MRI as input and low-count PET as the training target. We introduced a novel loss function that combines mean squared error (MSE) with a Bowsher prior (BP) to enforce structural consistency with the MRI. The method was evaluated on 29 ¹⁸F-FDG and 10 ¹⁸F-VAT PET/MRI scans using 10% Poisson-thinned PET data. Compared with Gaussian filtering, non-local means, N2V, and SSIM-based self-supervised alternatives, the proposed BP+MSE loss achieved superior PSNR, SSIM, and RMSE in ¹⁸F-FDG (p < 0.05). For ¹⁸F-VAT, BP+MSE showed the best mean performance across all metrics, though not achieving statistical significance when compared to N2V. Ablation studies confirmed the complementary value of combining BP and MSE and, demonstrated the benefit of the chosen network backbone. This flexible framework enables accurate recovery of full-count PET without requiring ground truth and generalizes across tracers.