Neural blind deconvolution for simultaneous partial volume correction and super-sampling of PSMA PET images

We aimed to simultaneously mitigate partial volume effects (PVEs) inprostate-specific membrane antigen (PSMA) positron emission tomography (PET)images while performing supersampling. Blind deconvolution is a method ofestimating the hypothetical "deblurred" image along with the blur kernel(related to the point spread function) simultaneously. Traditional maximum aposteriori blind deconvolution methods require stringent assumptions and sufferfrom convergence to a trivial solution. A promising method of modelling thedeblurred image and kernel with independent neural networks, called "neuralblind deconvolution" was demonstrated on 2D natural images in 2020. In thiswork, we adapt neural blind deconvolution for PVE correction of PSMA PETimages, along with simultaneous supersampling. We compare this methodology withseveral interpolation methods, using blind image quality metrics, and test themodel's ability to predict kernels by re-running the model after applyingartificial "pseudokernels" to deblurred images. Our results demonstrateimprovements in image quality over other interpolation methods in terms ofblind image quality metrics and visual assessment. Predicted kernels weresimilar between patients, and the model accurately predicted severalartificially-applied pseudokernels. The intrinsically low spatial resolution ofPSMA PET leads to PVEs which negatively impact uptake quantification in smallregions. The proposed method can be used to mitigate this issue, and can bestraightforwardly adapted for other medical imaging modalities.