Multi-Module Deep Learning for Enhanced and Accelerated PET Image Reconstruction

Neurological positron emission tomography (PET) provides a vital diagnostic and patient management tool for treating a wide range of conditions, such as oncological and neurodegenerative diseases. One particular research direction currently being explored to further the clinical utility of such scans is the inclusion of deep learning methodologies into PET image reconstruction algorithms. As such, this work presents an unrolled neural network architecture that combines the advantages of a known system matrix and projection operators with denoising filters trained by using a representative population database. The trade-off between model complexity and image reconstruction accuracy is investigated in addition to network architecture. The proposed method achieves superior results in comparison to the popular post-smoothed maximum likelihood expectation maximisation (MLEM) algorithm in terms of the image quality metric NRMSE, 10.6% vs 13.4% (for a given injected dose) and accelerates the speed of reconstruction 7.6s vs 16.8s. Further investigation is required to confirm these results using real patient datasets and exploring the added flexibility that the presented unrolled architecture achieves, in comparison to other alternative network architectures.