Effects of Target Images on Outputs of Style-Transfer Generative Models for Brain MRI Harmonization

Multi-site neuroimaging studies allow for increased sample sizes and improved statistical power for robustness and generalizability of findings. However, image acquisition protocols and scanner properties can contribute to site-specific effects, which may lead to biased results. Data harmonization can help mitigate scanner-induced biases, and recent generative approaches for transfering the style of images from one scanner to images from other scanners have been proposed. However, these approaches usually require a single representative image to serve as the target style, and the extent to which specific anatomical content in the image, such as abnormal pathology, may drive harmonization “style” has yet to be evaluated. Here, we compare MRI harmonization results from three style-transfer deep learning models: StarGANv2, ArtFlow and StyTr<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> to assess harmonization performance using a traveling subjects dataset and variability due to target selection using a dataset with varying levels of individual pathology in terms of white matter hyperintensities. We found StyTr<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> produced images where the harmonized results were less variant to white matter hyperintensity pathologies introduced in style. Our work suggests future harmonization methods may want to consider transformers such as StyTr<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> for more repeatable generated images.