Paired PET-MRI Deep Learning Model for Translating [¹¹C]PiB to [¹⁸F]Florbetaben Amyloid Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Amyloid PET imaging is crucial for assessing Alzheimer's disease (AD).
• Limited interchangeability among radiotracers complicates clinical trials and multicenter studies.
• The Centiloid method standardizes but lacks voxel-wise translation capabilities.

Purpose of the Study:

• To develop a DCNv3-based U-Net model for translating [¹¹C]-Pittsburgh compound B PET images to [¹⁸F]-florbetaben.
• To augment datasets for large-scale model training in Alzheimer's disease research.
• To provide consistent image information when visual assessments and Centiloid scale diverge.

Main Methods:

• Utilized a U-Net architecture incorporating deformable convolution (DCNv3).
• Leveraged DCNv3 for capturing long-range dependencies and efficient computation.
• Employed an encoder-decoder structure with skip connections for comprehensive feature learning and image synthesis.

Main Results:

• The DCNv3-based U-Net generated prediction images with enhanced homogeneity.
• The synthesized images closely matched the texture of [¹⁸F]-florbetaben PET scans.
• The model demonstrated high performance in quantitative metrics and statistical analyses.

Conclusions:

• The DCNv3-based U-Net effectively performs PET image-to-image translation.
• This approach shows promise for standardizing amyloid PET data in Alzheimer's disease research.
• Magnetic Resonance (MR) imaging is vital for providing essential structural information.