Three-Dimensional Amyloid-Beta PET Synthesis from Structural MRI with Conditional Generative Adversarial Networks

TL;DR

The model proves the feasibility of synthesizing amyloid-beta PET images from structural MRI ones, significantly enhancing accessibility for large-cohort studies and early dementia detection, while also reducing cost, invasiveness, and radiation exposure.

Abstract

Motivation: Alzheimer's Disease hallmarks include amyloid-beta deposits and brain atrophy, detectable via PET and MRI scans, respectively. PET is expensive, invasive and exposes patients to ionizing radiation. MRI is cheaper, non-invasive, and free from ionizing radiation but limited to measuring brain atrophy. Goal: To develop an 3D image translation model that synthesizes amyloid-beta PET images from T1-weighted MRI, exploiting the known relationship between amyloid-beta and brain atrophy. Approach: The model was trained on 616 PET/MRI pairs and validated with 264 pairs. Results: The model synthesized amyloid-beta PET images from T1-weighted MRI with high-degree of similarity showing high SSIM and PSNR metrics (SSIM>0.95&PSNR=28). Impact: Our model proves the feasibility of synthesizing amyloid-beta PET images from structural MRI ones, significantly enhancing accessibility for large-cohort studies and early dementia detection, while also reducing cost, invasiveness, and radiation exposure.

Figures (4)

• Figure 1: Proposed image translation model following a 3D-cGAN architecture. The generator uses an encoder-decoder architecture that receives an MRI input and generates a synthetic PET image (output) that is compared with the real PET (label). Then the real pair (real PET and MRI input) and synthetic pair (synthetic PET and MRI input) are independently fed into the discriminator which classifies the given pair as real or synthetic. The discriminator incorporated spectral normalization in the middle layers.
• Figure 2: Amyloid-beta synthesis comparison. Two subjects are compared, a cognitively normal (CN) and an Alzheimer’s Disease (AD) across three anatomical views: axial, coronal and sagittal. For each view an MRI, real PET, synthetic PET and a difference map between real and synthetic PET. Showing that the model can produce high-quality synthetic PET images for CN and AD cases that are close in shape and SUVR quantification.
• Figure 3: SSIM and PSNR data distribution across 264 cohort of unseen subjects. SSIM and its luminescence, contrast, and structure components are reported, reaching a mean SSIM above 0.95, mean luminescence component above 0.98, mean contrast component above 0.96 and mean structure component above 1. With a mean PSNR above 28 indicating that the model can produce synthetic PET images have high degree of similarity with the true ones.
• Figure 4: Detailed SUVR comparison between synthetic and real amyloid-beta PET images, showing the axial, coronal and sagittal planes. While the synthetic images highly resemble the real ones, there is still room for improvement. The AD case in the axial plane, shows that the tracer accuracy requires enhancements while the CN subject shows higher accuracy in the SUVR quantification.