Transfer learning improves supervised image segmentation across imaging protocols.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Variations in biomedical image acquisition (scanners, protocols) challenge automatic segmentation.
• Supervised learning methods require extensive, representative labeled data, which is often unavailable.
• Distribution shifts between training and target data limit supervised model generalization.

Purpose of the Study:

• To investigate transfer learning for biomedical image segmentation across different scanners and protocols.
• To develop and evaluate transfer classifiers capable of learning from limited representative data.
• To compare the performance of transfer learning against standard supervised learning in brain MRI segmentation.

Main Methods:

• Proposed four transfer learning classifiers designed for limited representative training data.
• Utilized multi-site magnetic resonance imaging (MRI) brain datasets.
• Compared transfer classifiers against standard supervised classification on two segmentation tasks: tissue segmentation (white matter, gray matter, CSF) and lesion segmentation (white matter/MS lesions).

Main Results:

• Transfer learning significantly outperformed supervised learning when only a small amount of representative training data was available.
• Classification errors were minimized by up to 60% using transfer learning approaches.
• Demonstrated the effectiveness of transfer learning in handling distribution differences in medical image segmentation.

Conclusions:

• Transfer learning is a robust solution for biomedical image segmentation challenges caused by data variability.
• The proposed transfer classifiers effectively leverage limited representative data for improved segmentation performance.
• Transfer learning offers a promising alternative to supervised learning for cross-scanner and cross-protocol medical image segmentation.