Unsupervised Domain Adaptation of Brain MRI Skull Stripping Trained on Adult Data to Newborns: Combining Synthetic Data with Domain Invariant Features

Abbas Omidi, Amirmohammad Shamaei, Anouk Verschuu, Regan King, Lara Leijser, Roberto Souza
Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, PMLR 250:1073-1085, 2024.

Abstract

Skull-stripping constitutes a crucial initial step in neuroimaging analysis, and supervised deep-learning models have demonstrated considerable success in automating this task. However, a notable challenge is the limited availability of publicly accessible newborn brain MRI datasets. Furthermore, these datasets frequently use diverse post-processing techniques to improve image quality, which may not be consistently feasible in all clinical settings. Additionally, manual segmentation of newborn brain MR images is labor-intensive and demands specialized expertise, rendering it inefficient. While various adult brain MRI datasets with skull-stripping masks are publicly available, applying supervised models trained on these datasets directly to newborns poses a challenge due to domain shift. We propose a methodology that combines domain adversarial models to learn domain-invariant features between newborn and adult data, along with the integration of synthetic data generated using a Gaussian Mixture Model (GMM) as well as data augmentation procedures. The GMM method facilitates the creation of synthetic brain MR images, ensuring a diverse and representative input from multiple domains within our source dataset during model training. The data augmentation procedures were tailored to make the adult MRI data distribution closer to the newborn data distribution. Our results yielded an overall Dice coefficient of 0.9308 ± 0.0297 (mean± std), outperforming all compared unsupervised domain adaptation models and surpassing some supervised techniques previously trained on newborn data. This projectś code and trained models\’{weights} are publicly available at https://github.com/abbasomidi77/GMM-Enhanced-DAUnet

Cite this Paper

BibTeX
@InProceedings{pmlr-v250-omidi24a, title = {Unsupervised Domain Adaptation of Brain MRI Skull Stripping Trained on Adult Data to Newborns: Combining Synthetic Data with Domain Invariant Features}, author = {Omidi, Abbas and Shamaei, Amirmohammad and Verschuu, Anouk and King, Regan and Leijser, Lara and Souza, Roberto}, booktitle = {Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning}, pages = {1073--1085}, year = {2024}, editor = {Burgos, Ninon and Petitjean, Caroline and Vakalopoulou, Maria and Christodoulidis, Stergios and Coupe, Pierrick and Delingette, Hervé and Lartizien, Carole and Mateus, Diana}, volume = {250}, series = {Proceedings of Machine Learning Research}, month = {03--05 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v250/main/assets/omidi24a/omidi24a.pdf}, url = {https://proceedings.mlr.press/v250/omidi24a.html}, abstract = {Skull-stripping constitutes a crucial initial step in neuroimaging analysis, and supervised deep-learning models have demonstrated considerable success in automating this task. However, a notable challenge is the limited availability of publicly accessible newborn brain MRI datasets. Furthermore, these datasets frequently use diverse post-processing techniques to improve image quality, which may not be consistently feasible in all clinical settings. Additionally, manual segmentation of newborn brain MR images is labor-intensive and demands specialized expertise, rendering it inefficient. While various adult brain MRI datasets with skull-stripping masks are publicly available, applying supervised models trained on these datasets directly to newborns poses a challenge due to domain shift. We propose a methodology that combines domain adversarial models to learn domain-invariant features between newborn and adult data, along with the integration of synthetic data generated using a Gaussian Mixture Model (GMM) as well as data augmentation procedures. The GMM method facilitates the creation of synthetic brain MR images, ensuring a diverse and representative input from multiple domains within our source dataset during model training. The data augmentation procedures were tailored to make the adult MRI data distribution closer to the newborn data distribution. Our results yielded an overall Dice coefficient of 0.9308 ± 0.0297 (mean± std), outperforming all compared unsupervised domain adaptation models and surpassing some supervised techniques previously trained on newborn data. This projectś code and trained models\’{weights} are publicly available at https://github.com/abbasomidi77/GMM-Enhanced-DAUnet} }
Endnote
%0 Conference Paper %T Unsupervised Domain Adaptation of Brain MRI Skull Stripping Trained on Adult Data to Newborns: Combining Synthetic Data with Domain Invariant Features %A Abbas Omidi %A Amirmohammad Shamaei %A Anouk Verschuu %A Regan King %A Lara Leijser %A Roberto Souza %B Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2024 %E Ninon Burgos %E Caroline Petitjean %E Maria Vakalopoulou %E Stergios Christodoulidis %E Pierrick Coupe %E Hervé Delingette %E Carole Lartizien %E Diana Mateus %F pmlr-v250-omidi24a %I PMLR %P 1073--1085 %U https://proceedings.mlr.press/v250/omidi24a.html %V 250 %X Skull-stripping constitutes a crucial initial step in neuroimaging analysis, and supervised deep-learning models have demonstrated considerable success in automating this task. However, a notable challenge is the limited availability of publicly accessible newborn brain MRI datasets. Furthermore, these datasets frequently use diverse post-processing techniques to improve image quality, which may not be consistently feasible in all clinical settings. Additionally, manual segmentation of newborn brain MR images is labor-intensive and demands specialized expertise, rendering it inefficient. While various adult brain MRI datasets with skull-stripping masks are publicly available, applying supervised models trained on these datasets directly to newborns poses a challenge due to domain shift. We propose a methodology that combines domain adversarial models to learn domain-invariant features between newborn and adult data, along with the integration of synthetic data generated using a Gaussian Mixture Model (GMM) as well as data augmentation procedures. The GMM method facilitates the creation of synthetic brain MR images, ensuring a diverse and representative input from multiple domains within our source dataset during model training. The data augmentation procedures were tailored to make the adult MRI data distribution closer to the newborn data distribution. Our results yielded an overall Dice coefficient of 0.9308 ± 0.0297 (mean± std), outperforming all compared unsupervised domain adaptation models and surpassing some supervised techniques previously trained on newborn data. This projectś code and trained models\’{weights} are publicly available at https://github.com/abbasomidi77/GMM-Enhanced-DAUnet
APA
Omidi, A., Shamaei, A., Verschuu, A., King, R., Leijser, L. & Souza, R.. (2024). Unsupervised Domain Adaptation of Brain MRI Skull Stripping Trained on Adult Data to Newborns: Combining Synthetic Data with Domain Invariant Features. Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 250:1073-1085 Available from https://proceedings.mlr.press/v250/omidi24a.html.

Related Material