Convolutional micro-networks for MR-guided low-count PET image processing

Lowering patient radiation dose or acquisition times in positron emission tomography (PET) results in low count data and reconstructed images with low signal to noise ratio (SNR). Anatomical information can be incorporated from jointly acquired modalities such as computed tomography (CT) or magnetic resonance (MR) images to improve PET image quality. Convolutional neural networks (CNNs) are particularly well suited to joint image processing, but have mostly been applied outside the field of medical imaging or used for classification and segmentation rather than image quality improvement. We propose a deliberately small network (µ-net) which can be trained with comparatively little data (two simulated noise realisations of a single phantom). The network is trained to post-process reconstructed images and is demonstrated to reduce normalised root mean squared error (NRMSE) by between 45 % to 80 % compared to maximum likelihood expectation maximisation (MLEM) for low (21.5 M) and very low (4.3 M) count validation data simulations. Furthermore, a µ-net prediction based on a real low (43 M) count scan is also demonstrated to be comparable to a full count reconstruction, suppressing noise and recovering resolution without introducing ringing artefacts.