ESTRO 2021 Abstract Book

S411

ESTRO 2021

Results We report (Table 1) categorical accuracy on the source scanner ID of harmonised images, classified with an independent CNN trained on the full 6-scanner input dataset and validated on test data. Harmonised ID accuracy (mean = 5.3%) represents the rate at which source scanner could be classified from harmonised images, with zero representing perfect harmonisation. Harmonisation from GE-1.5T to GE-3T was poorest, suggesting that whilst it is relatively easy to remove manufacturer dependence, it is very challenging to generate high-field images from low field ones. Anatomical detail is well preserved (fig. 1b). Normalised mutual-information (NMI) between input and harmonised images revealed strong content preservation (mean NMI=0.88) following harmonisation.

Conclusion HarMonAE was able to harmonise unpaired T1w MR from multiple source scanners, removing scanner-specific information whilst preserving anatomical information. Zero-shot performance on images from an unseen scanner was comparable to that for in-training source scanners, significantly broadening future clinical applicability. Harmonisation of MR enhances and enables many radiomics and deep-learning solutions in