Unsupervised and Weakly Supervised Domain Adaptation of MRI Skull-Stripping Models Trained on Adult Data to Newborns
Date
2025-01-13
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The process of removing non-brain tissue signals from brain magnetic resonance imaging (MRI) is known as skull-stripping. It is a crucial preprocessing step in neuroimaging analysis, particularly for subsequent brain tissue segmentation and studying neurological disorders. Despite significant progress in deep learning-based methods for skull-stripping, data distribution shifts between adult and newborn MRI data present a major challenge, limiting the generalization of models trained on adult data when applied to newborns. This work proposes both unsupervised and weakly supervised domain adaptation techniques that leverage weakly annotated data, synthetic data, and the learning of domain-invariant features to address this challenge. First, I introduce an unsupervised method utilizing adversarial domain adaptation to align feature representations between adult and newborn MRI data, and a new contrast inversion data augmentation step to reduce the domain shift. Then, I extend this method by leveraging Gaussian Mixture Model (GMM)-generated synthetic data to enhance segmentation performance. Finally, I propose to incorporate weakly annotated newborn data during model training. This weakly supervised method achieves state-of-the-art performance for skull-stripping neonatal brain imaging, improving upon existing methods in terms of both the Dice coefficient and Hausdorff distance quantitative metrics. Together, these methods demonstrate the potential of leveraging domain adaptation techniques to bridge the gap between adult and newborn brain MRI data, enabling accurate skull-stripping across diverse populations. The source code and weights of the trained models are publicly available at https://github.com/abbasomidi77/DAUnet.
Description
Keywords
Skull-stripping, Deep Learning, Artificial Intelligence, Machine Learning, Computer Vision, Magnetic Resonance Imaging, Synthetic Data, Infants, Domain Adaptation
Citation
Omidi, A. (2025). Unsupervised and weakly supervised domain adaptation of MRI skull-stripping models trained on adult data to newborns (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.