Self-Supervised Early Stopping for End-To-End PET Reconstruction With Implicit Regularization Through Neural-Network Architectures

Résumé fourni par la source

Neural networks (NN) can serve as implicit regularizers in unsupervised PET image reconstruction by encoding prior information through their architecture. However, such methods are prone to overfitting if training continues too long, leading to noise amplification. In this work, we study the variability of end-to-end PET reconstruction using Deep Image Prior and SIREN architectures, and evaluate a self-supervised early stopping (ES) strategy based on the windowed moving variance (WMV) of the reconstruction sequence. We design experiments to isolate the effects of weight initialization, random input, and PET simulation variability, and demonstrate that the WMV-based ES method yields reconstructions close to the optimal stopping point. Our results show that self-supervised ES is feasible and enables implicit regularization through NNs without being altered by using explicit priors.