SURE-based Stopping Strategy for Fine-tunable Supervised PET Image Denoising

Previously, based on the unsupervised deep learning method, conditional deep image prior (CDIP), we showed the possibility of finetuning during the testing phase after supervised learning. However, one barrier for CDIP finetuning is that it needs early stopping before the network output converges to the noise which is hard to determine. In this work, we proposed to construct the finetuning loss based on Stein’s unbiased risk estimator (SURE), which can provide an unbiased estimation of the true mean square error (MSE). We demonstrated that using SURE-based finetuning loss, the network output can be close to the ground truth while not converging to the noise. Quantification results based on seven real patient datasets show that the proposed method is more stable during finetuning than CDIP.