HYPR4D Kernel Method With an Unsupervised 2.5SD+0.5TD Deep Learning Assisted Kernel Matrix

Abstract

We describe a deep learning (DL) assisted HYPR4D kernelized reconstruction which produces low-noise voxel-level time-activity-curves (TACs) while preserving quantification within small structures as well as consistent spatiotemporal patterns/features within measured data. The proposed method consists of the following advantages over other methods: 1) unsupervised single subject network training scheme independent of positron emission tomography (PET) tracers; 2) training data generated on-the-fly during reconstruction; 3) intrinsic spatiotemporal consistency provided by minimizing the $L_{2}$ loss using pseudo 4-D (i.e., 2.5 Spatial Dimension + 0.5 Temporal Dimension or 2.5SD+0.5TD) patches between kernelized OSEM subset estimates; and 4) a final tuning step which minimizes over-smoothing from the network output within the kernel matrix. Contrast phantom, human [18F]FDG and [11C]RAC data acquired on GE SIGNA PET/MR were used for evaluations. The proposed DL HYPR4D kernel method outperformed the standard HYPR4D kernel method as well as TOF-OSEM and TOF-BSREM (Q.Clear) in terms contrast recovery versus noise. The proposed final tuning reduced the underestimation bias due to over-smoothing within a 4-mm target structure from ~15% to ~2% while maintaining low-noise voxel-level TACs. In addition, the proposed unsupervised DL assisted reconstruction also outperformed the supervised DL version in terms of bias reduction along the TACs and kinetic model fittings.