Tau PET Image Harmonization Using a Generative Adversarial Network
Amirhossein Behzadfar, Tzu-An Song, Cristina Lois Gómez, Kyungsang Kim, Gad A. Marshall, Keith A. Johnson, Joyita Dutta · 2022
Multi-scanner image harmonization is a major challenge for longitudinal imaging studies that often span several decades and transition from legacy scanners to state-of-the-art alternatives. In this preliminary tau PET imaging study involving 5 human subjects sequentially scanned on a GE Discovery MI PET/CT scanner and a Siemens HR+ scanner, we train and validate a generative adversarial network (GAN) for harmonizing tau PET images between the two scanners. Our network receives as inputs an HR+ image, a high-resolution anatomical MR image, and spatial information. The method was validated on clinical data by comparing the synthetic images generated by the GAN with those generated by a convolutional neural network (CNN) and by deconvolution stabilized by a total variation penalty. Our results show that the GAN outperforms both classical deblurring and the CNN in terms of image quality and quantitation as indicated by the peak signal-to-noise-ratio.