Abstract: We propose a novel micro-network (mu-net) architecture which is a convolutional neural network (CNN) designed to be robust to minimal training data. Applying the mu-net to MR-guided low count PET scans produces impressive qualitative and quantitative cross-validation results despite training on just two noise realisations of the same patient.
Author Affiliation: King's College London
Keywords: CNN, image processing, guided reconstruction, PET-MR
3 Replies
Loading