Convolutional Neural Network Application for Brain PET Image Translation to Support Alzheimer's Disease Diagnosis

Alzheimer's Disease (AD) is supported by several complementary neuroimaging scans including 18F - fluorodeoxyglucose PET (FDG-PET),
¹⁸
F-AV45 florbetapir PET (AV45-PET) and tau PET. Limiting PET scans would allow reducing cost and radiation exposure. Interestingly, the early PET acquisition of the AV45 scan would allow providing information similar to a FDG-PET scan. This study aims to perform an unidirectional translation from eAV45-PET to FDG-PET to optimize the similarities between both scan so that both complementary informations could be obtained from a single visit and radiotracer injection. The eAV45-PET to FDG-PET image translation is made by a ResN et neural network as well as a specific encoding and decoding algorithms. Results show good performances (0.00160 for Mean Square Error, 0.10301 for Normalized Root Mean Square Error, 33.80288 for Peak Signal to Noise Ratio and 33.80288 for Structural Similarity), however the technique still needs to be optimized to reduce the intensity discrepancy in some brain regions especially the frontal cortex.