Domain Unlearning Boosts Lesion Segmentation Performance on Seen and Unseen MR Scanner Data

Inter-scanner variability was highlighted in the 2020 MAGNIMS consensus guidelines as most-detrimental factor in image acquisition with high impact on the diagnostic and prognostic quality of the MR scans. This study aimed to evaluate and compare the contributions of domain unlearning on white matter lesion segmentation from the MR scans. We used the MSSEG 2016 challenge dataset of 53 MS patients, where MR images were acquired on 4 different scanners, one of which was absent in the training split. Our approach was the state-of-the-art Swin UNETR segmentation model that was adopted for domain unlearning to better handle inter-scanner bias. Performance of lesion segmentation was evaluated according to challenge protocol and compared to three best challenge results and the state-of-the-art results of recently proposed Dense Residual UNET method. The baseline Swin UNETR model achieved comparable results to the three challenge methods, while our domain unlearning model consistently improved in all metrics versus the baseline model and achieved the highest Dice score among all tested methods. The principal impact of unlearning was in reducing false positive annotations, whereas performance consistently improved on seen and unseen scanner data, indicating that scanner-specific intensity artifacts do not overlap with the information required for lesion segmentation.