Regional (18)F-fluoromisonidazole PET images generated from multiple advanced MR images using neural networks in glioblastoma

Generated (18)F-fluoromisonidazole ((18)F-FMISO) positron emission tomography (PET) images for glioblastoma are highly sought after because (18)F-FMISO can be radioactive, and the imaging procedure is not easy. This study aimed to explore the feasibility of using advanced magnetic resonance (MR) images to generate regional (18)F-FMISO PET images and its predictive value for survival. Twelve kinds of advanced MR images of 28 patients from The Cancer Imaging Archive were processed. Voxel-by-voxel correlation analysis between (18)F-FMISO images and advanced MR images was performed to select the MR images for generating regional (18)F-FMISO images. Neural network algorithms provided by the MATLAB toolbox were used to generate regional (18)F-FMISO images. The mean square error (MSE) was used to evaluate the regression effect. The prognostic value of generated (18)F-FMISO images was evaluated by the Mantel-Cox test. A total of 299 831 voxels were extracted from the segmented regions of all patients. Eleven kinds of advanced MR images were selected to generate (18)F-FMISO images. The best neural network algorithm was Bayesian regularization. The MSEs of the training, validation, and testing groups were 2.92E-2, 2.9E-2, and 2.92E-2, respectively. Both the maximum Tissue/Blood ratio (P = .017) and hypoxic volume (P = .023) of the generated images were predictive factors of overall survival, but only hypoxic volume (P = .029) was a predictive factor of progression-free survival. Multiple advanced MR images are feasible to generate qualified regional (18)F-FMISO PET images using neural networks. The generated images also have predictive value in the prognostic evaluation of glioblastoma.