Our paper Evaluating Deep Learning Based Domain Generalization for Motion Mitigation in Multi-Center Brain MRI has been accepted to MICCAI-RIME 2025 workshop.

Paper Title: Evaluating Deep Learning Based Domain Generalization for Motion Mitigation in Multi-Center Brain MRI

Authors: Saad Bin Ashraf, Md Afif Al Mamun, Mumu Akter, Roberto Souza, Mariana Pinheiro Bento

Abstract

Magnetic Resonance Imaging (MRI) is an essential tool for diagnosing brain conditions. The scan procedure is time-consuming, during which the patient must remain still. Any movement during the scan can cause motion artifacts, appearing as noise artifacts in the reconstructed image, complicating diagnosis. Recent Deep Learning (DL) models are effective in tasks such as skull stripping, tissue segmentation, and motion mitigation. However, DL models struggle with distribution shifts occurring due to MRI scans collected in different centers, making it harder to adapt to different datasets. Additionally, unlike other tasks, motion mitigation works with noisy MRI scans, which are harder to denoise since the original scan is distorted. Most of the motion mitigation models have been trained on single datasets; however, for usage in real life, it is crucial to explore the domain adaptability of such models. In our study, we have used three open datasets, collected from 11 different centers. Our chosen baseline 3D UNet model was trained on individual datasets and also in different combinations of these datasets. The model trained on a large, diverse dataset could preserve its knowledge compared to the current literature, while models trained on smaller datasets performed better on datasets with similar properties to the training dataset. These insights can be used to drive further data preprocessing techniques for domain adaptation research concerning motion-mitigation.