Unsupervised domain adaptation for medical imaging segmentation with self-ensembling

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning excels in medical imaging but often fails to generalize across different datasets or domains.
• Variability in medical images, even from the same modality like MRI, poses a significant challenge for model generalizability.

Purpose of the Study:

• To extend unsupervised domain adaptation using self-ensembling for semantic segmentation in medical imaging.
• To evaluate the effectiveness of self-ensembling on a small, realistic MRI dataset.

Main Methods:

• Applied unsupervised domain adaptation with self-ensembling to semantic segmentation tasks.
• Conducted extensive evaluations on a publicly available magnetic resonance imaging (MRI) dataset.

Main Results:

• Demonstrated that self-ensembling enhances model generalization capabilities.
• Showed improved performance even when utilizing a small quantity of unlabeled data.

Conclusions:

• Self-ensembling is a viable technique to improve the generalizability of deep learning models in medical imaging.
• The method shows promise for overcoming domain shift challenges in applications like MRI semantic segmentation.