Multi-Source Domain Generalization for Learned Lossless Volumetric Biomedical Image Compression.

Area of Science:

• Medical Imaging
• Computer Vision
• Data Compression

Background:

• Learned lossless compression significantly improves volumetric biomedical image compression over traditional methods.
• Domain gap issues, arising from modality and structure differences, cause poor performance on unseen data.
• Existing methods struggle with generalization across diverse biomedical imaging datasets.

Purpose of the Study:

• To develop a multi-source domain generalization method for volumetric biomedical image compression.
• To address both modality and structure differences causing domain gaps.
• To improve the robustness and applicability of learned compression techniques in unseen biomedical imaging domains.

Main Methods:

• Proposed an adaptive modality transfer (AMT) module using dynamic convolutions controlled by modality-specific parameters embedded in the bitstream.
• Designed an adaptive structure transfer (AST) module decomposing images into LSB and MSB in the wavelet domain for structure adaptation.
• Integrated AMT and AST to enable multi-source domain generalization for compression.

Main Results:

• Achieved performance degradation within 3% across various volumetric biomedical modalities despite domain gaps.
• Demonstrated effective handling of both modality and structure differences.
• Showcased significant improvement in generalization capabilities for learned compression methods.

Conclusions:

• The proposed multi-source domain generalization method effectively bridges the domain gap in volumetric biomedical image compression.
• This approach enables practical, end-to-end compression solutions for diverse biomedical imaging applications.
• Significant reduction in performance loss makes learned compression more viable for real-world medical imaging scenarios.