Amyloid‐Beta PET Synthesis from Structural MRI: A Potential Alternative Method for Alzheimer’s Disease Screening

Abstract

AbstractBackgroundAmyloid‐beta and structural brain atrophy are known to be hallmarks of Alzheimer’s Disease (AD), and can be quantified with Positron Emission Tomography (PET) and structural Magnetic Resonance (MRI), respectively. PET scans use radiotracers that binds to amyloid‐beta molecules, whereas MRI measures changes in structural morphology. PET scans are difficult to perform due to cost (∼$5000/scan), invasiveness, and ionizing radiation exposure, making them inaccessible for screening early‐onset AD. Conversely, MRI is a cheaper (∼$500/scan), non‐invasive, and free from ionizing radiation technique, however, it cannot provide molecular information obtained from PET. There is a known relationship between amyloid‐beta and brain atrophy, hence, amyloid‐beta PET images might be synthesized from structural MRI using image translation, which is an advanced form of machine learning.MethodAn image translation algorithm was developed using the Open Access Series of Imaging Studies (OASIS‐3) dataset that provides 929 subjects with pairs of T1‐weighted MRI and amyloid‐beta PET images, where 609 are cognitively normal (CN) and 489 at different stages of cognitive decline based on Clinical Dementia Rating scores. The image translation algorithm implemented with Conditional Generative Adversarial Networks, shown in Figure 1, was trained with 550 pairs with 156 axial slices each of amyloid‐beta and MRI images. Eight models where generated with different parameters to assess their impact in the generalization quality, then compared in a sample of 334 subjects using Structural Similarity Index (SSIM) and Peak Signal‐to‐Noise Ratio (PSNR).ResultFigure 2, compares the models performance based on SSIM and PSNR, where the best performing model generates synthetic PET images with a high degree of similarity with the real ones in Figure 3.ConclusionAn image translation pipeline was implemented and explored by changing parameters that can affect the model performance, the developed models are robust against the different parameters that were changed and it demonstrates that amyloid‐beta PET images can be synthesized from structural MRI, reducing cost, invasiveness and increasing availability. Such a method could enable non‐invasive screening of early onset AD.