Implementing the Point Spread Function Deconvolution for Better Molecular Characterization of Newly Diagnosed Gliomas: A Dynamic (18)F-FDOPA PET Radiomics Study

SIMPLE SUMMARY: The point spread function deconvolution (PSFd), which is known to improve contrast and spatial resolution of brain positron emission tomography (PET) images, has not been evaluated for the routine analysis of amino-acid PET imaging. Our study therefore investigated the effects of applying the PSFd to a radiomics analysis of the clinical dynamic L-3,4-dihydroxy-6-[(18)F]-fluoro-phenyl-alanine ((18)F-FDOPA) PET images (tumor-to-background ratio and time-to-peak parametric images), and evaluated the impact of these effects on the molecular characterization of newly diagnosed gliomas. We show that applying the PSFd to dynamic (18)F-FDOPA PET images significantly improves the detection of molecular parameters in newly diagnosed gliomas for predicting isocitrate dehydrogenase mutated and/or 1p/19q codeleted gliomas, for a combination of radiomics features extracted from static and dynamic parametric images. ABSTRACT: Purpose: This study aims to investigate the effects of applying the point spread function deconvolution (PSFd) to the radiomics analysis of dynamic L-3,4-dihydroxy-6-[(18)F]-fluoro-phenyl-alanine ((18)F-FDOPA) positron emission tomography (PET) images, to non-invasively identify isocitrate dehydrogenase (IDH) mutated and/or 1p/19q codeleted gliomas. Methods: Fifty-seven newly diagnosed glioma patients underwent dynamic (18)F-FDOPA imaging on the same digital PET system. All images were reconstructed with and without PSFd. An L1-penalized (Lasso) logistic regression model, with 5-fold cross-validation and 20 repetitions, was trained with radiomics features extracted from the static tumor-to-background-ratio (TBR) and dynamic time-to-peak (TTP) parametric images, as well as a combination of both. Feature importance was assessed using Shapley additive explanation values. Results: The PSFd significantly modified 95% of TBR, but only 79% of TTP radiomics features. Applying the PSFd significantly improved the ability to identify IDH-mutated and/or 1p/19q codeleted gliomas, compared to PET images not processed with PSFd, with respective areas under the curve of 0.83 versus 0.79 and 0.75 versus 0.68 for a combination of static and dynamic radiomics features (p < 0.001). Without the PSFd, four and eight radiomics features contributed to 50% of the model for detecting IDH-mutated and/or 1p/19q codeleted gliomas, respectively. Application of the PSFd reduced this to three and seven contributive radiomics features. Conclusion: Application of the PSFd to dynamic (18)F-FDOPA PET imaging significantly improves the detection of molecular parameters in newly diagnosed gliomas, most notably by modifying TBR radiomics features.