DALSA: Domain Adaptation for Supervised Learning From Sparsely Annotated MR Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Supervised learning for automated tissue classification requires extensive manual segmentation.
• Manual annotation is labor-intensive, complex, and prone to errors, hindering practical application.
• Existing methods need retraining for each new application, site, or acquisition setup.

Purpose of the Study:

• To develop a method that effectively corrects sampling selection errors from sparse annotations in supervised learning.
• To enable high-quality tissue classifiers using sparse, unambiguous annotations.
• To reduce the dependency on large, fully annotated datasets for automated tumor segmentation.

Main Methods:

• Employing transfer learning techniques for error correction.
• Utilizing domain adaptation for sparse sampling.
• Validating the approach on multi-modal MR images of malignant gliomas and BraTS 2013 data.

Main Results:

• Reduced labeling time by over 70x and training time by over 180x compared to fully labeled data.
• Achieved accuracy comparable to methods trained on fully labeled datasets.
• Demonstrated the method's effectiveness in correcting sampling selection errors.

Conclusions:

• The proposed method significantly eases the creation and extension of annotated databases.
• This approach is a crucial step towards the practical application of learning-based tissue classification.
• Enables more efficient and accurate automated tumor segmentation in diverse clinical settings.