Neural blind deconvolution for deblurring and supersampling PSMA PET

Abstract

Objective: To simultaneously deblur and supersample prostate specific membrane antigen (PSMA) positron emission tomography (PET) images using neural blind deconvolution. Approach: Blind deconvolution is a method of estimating the hypothetical "deblurred" image along with the blur kernel (related to the point spread function) simultaneously. Traditional maximum a posteriori blind deconvolution methods require stringent assumptions and suffer from convergence to a trivial solution. A method of modelling the deblurred image and kernel with independent neural networks, called "neural blind deconvolution" had demonstrated success for deblurring 2D natural images in 2020. In this work, we adapt neural blind deconvolution for PVE correction of PSMA PET images with simultaneous supersampling. We compare this methodology with several interpolation methods, using blind image quality metrics, and test the model's ability to predict kernels by re-running the model after applying artificial "pseudokernels" to deblurred images. The methodology was tested on a retrospective set of 30 prostate patients as well as phantom images containing spherical lesions of various volumes. Main Results: Neural blind deconvolution led to improvements in image quality over other interpolation methods in terms of blind image quality metrics, recovery coefficients, and visual assessment. Predicted kernels were similar between patients, and the model accurately predicted several artificially-applied pseudokernels. Localization of activity in phantom spheres was improved after deblurring, allowing small lesions to be more accurately defined. Significance: The intrinsically low spatial resolution of PSMA PET leads to PVEs which negatively impact uptake quantification in small regions. The proposed method can be used to mitigate this issue, and can be straightforwardly adapted for other imaging modalities.

Image denoising and model-independent parameterization for IVIM MRI

“…Predictions were mostly confined to within one voxel of the kernel centres. Similar truncation was reported when deblurring PSMA PET images (Sample et al 2023); however, voxels were more truncated in the present study. This may be partially explained by the smaller kernel size employed in this study.…”

“…Neural blind deconvolution was previously adapted for 3D prostate specific membrane antigen (PSMA) positron emission tomography (PET) in 2023, while incorporating simultaneous super-sampling into the methodology (Sample et al 2023). It was shown to improve blind image quality metrics, and strengthen correlations between PSMA PET uptake and sub-regional importance estimates in the parotid gland for predicting post-radiotherapy xerostomia (subjective dry mouth) (Sample et al 2024).…”

“…To evaluate the model's ability to predict accurate blur kernels, we employed the same strategy used to evaluate PSMA PET image deblurring (Sample et al 2023). This involves generating artificial kernels, convolving them with previously denoised images, then restarting neural blind deconvolution.…”

“…This may be partially explained by the smaller kernel size employed in this study. Considering that a relatively small kernel for IVIM MRI images was anticipated (due to generally sharper MR images than PET) while also considering memory constraints, we decreased the kernel size to 11 × 11 × 11 from the 15 × 15 × 15 size used for PSMA PET (Sample et al 2023(Sample et al , 2024. Furthermore, the relatively high noise present in EPI-based IVIM MRI is expected to complicate blur kernel estimation.…”

“…A multi-scale optimization procedure was implemented, as recommended for neural blind deconvolution by Kotera et al (2021) and recently implemented for PET denoising (Sample et al 2023(Sample et al , 2024. This involved pretraining networks using images that were first down-sampled by a factor of two, then up-sampling final outputs by 2 before performing a second round of pre-training.…”