Unsupervised Domain Adaptation of MRI Skull-stripping Trained on Adult Data to Newborns
Abdollah Omidi, Aida Mohammadshahi, Neha Gianchandani, Regan King, Lara M. Leijser, Roberto Souza
Abstract
Skull-stripping is an important first step when analyzing brain Magnetic Resonance Imaging (MRI) data. Deep learning-based supervised segmentation models, such as the U-net model, have shown promising results in automating this segmentation task. However, when it comes to newborn MRI data, there are no publicly available brain MRI datasets that come with manually annotated segmentation masks to be used as labels during the training of these models. Manual segmentation of brain MR images is time-consuming, labor-intensive, and requires expertise. Furthermore, using a segmentation model trained on adult brain MR images for segmenting newborn brain images is not effective due to a large domain shift between adult and newborn data. As a result, there is a need for more efficient and accurate skull-stripping methods for newborns’ brain MRIs. In this paper, we present an unsupervised approach to adapt a U-net skull-stripping model trained on adult MRI to work effectively on newborns. Our results demonstrate the effectiveness of our novel unsupervised approach in enhancing segmentation accuracy. Our proposed method achieved an overall Dice coefficient of 0.916 ± 0.032 (mean ± std), and our ablation studies confirmed the effectiveness of our proposal. Remarkably, despite being unsupervised, our model’s performance stands in close proximity to that of the current state-of-the-art supervised models against which we conducted our comparisons. These findings indicate the potential of this method as a valuable, easier, and faster tool for supporting healthcare professionals in the examination of MR images of newborn brains. All the codes are available at: https://github.com/abbasomidi77/DAUnet.