MIMRTL researchers and collaborators, led by S. Iman Zare Estakhraji, PhD recently published a paper that explored the use of unsupervised and supervised training methods to generate synthetic computed tomography (sCT) images from magnetic resonance (MR) images. The study employed a cycleGAN method with unpaired data sets for unsupervised training and several supervised models that were fine-tuned or trained from scratch. Co-authors included Ali Pirasteh, MD; Tyler Bradshaw, PhD; and Alan McMillan, PhD.
Results revealed that supervised training yielded significantly higher quantitative accuracy in generating sCT images, while fine-tuned training showed no clear advantage over supervised learning. The findings emphasize the importance of well-registered paired data sets for training compared to a large set of unpaired data. In summary, a sufficient amount of high quality input data is more important than merely a large quantity of input data.
Read the full paper here: https://doi.org/10.1016/j.compmedimag.2023.102227