Transfer Learning for Image Segmentation by Combining Image Weighting and Kernel Learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Supervised voxel classification is common for medical image segmentation.
• Performance degrades when training and test data distributions differ (e.g., scanner variations).
• Existing methods improve performance by weighting training data but don't alter data distributions.

Purpose of the Study:

• Investigate kernel learning to reduce training-test data distribution differences.
• Explore kernel learning's added value for medical image weighting.
• Propose a novel image weighting method using maximum mean discrepancy (MMD).

Main Methods:

• Kernel learning applied to reduce data distribution discrepancies.
• Image weighting strategies, including a new MMD-based approach.
• Joint optimization of image weights and kernel parameters.
• Experimental validation on brain tissue, white matter lesion, and hippocampus segmentation.

Main Results:

• Both kernel learning and image weighting significantly enhance segmentation performance on heterogeneous data.
• The proposed MMD weighting method achieves performance comparable to existing weighting techniques.
• Combining kernel learning and image weighting offers marginal additional performance improvements.

Conclusions:

• Kernel learning and image weighting are effective strategies for improving medical image segmentation robustness.
• The novel MMD-based weighting method provides a viable alternative for data distribution alignment.
• Joint optimization of kernel learning and image weighting can yield synergistic performance benefits.