Unsupervised PET Image Reconstruction using Deep Diffusion Bridge Image Prior

Deep learning-based PET reconstruction has shown promise but often requires tracer-specific training, limiting generalizability. Unsupervised approaches like Deep Image Prior (DIP) avoid the need for training data but typically underperform conventional regularized methods, especially when anatomical priors (e.g., MRI) are used. Moreover, conventional methods require manual parameter tuning, which can be suboptimal.