Learning Generalized Medical Image Representation by Decoupled Feature Queries.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical images are often acquired from diverse clinical centers using various scanner types, leading to significant cross-domain distribution discrepancies.
• Deep learning models can exhibit channel redundancy, where similar patterns are captured by multiple channels and different cross-domain patterns coexist within the same channel.
• This redundancy limits the expressive power of learned representations, hindering generalization ability in medical image analysis.

Purpose of the Study:

• To propose a novel Decoupled Feature as Query (DFQ) framework for domain generalized medical image representation learning.
• To address the challenge of channel redundancy in deep networks when learning from multi-center, multi-scanner medical image data.
• To enhance the generalization capability of medical image analysis models across different domains.

Main Methods:

• The proposed Decoupled Feature as Query (DFQ) framework leverages channel-wise decoupled deep features as queries.
• A deep instance whitening transform with restricted isometry is introduced to enforce orthogonality between decoupled channels.
• Long-range dependencies between decoupled deep and shallow features are implicitly constrained to minimize channel redundancy during training.

Main Results:

• The DFQ framework demonstrates state-of-the-art performance on domain generalization tasks in medical imaging.
• Experiments were conducted on three distinct medical domain generalization tasks.
• The framework was evaluated using four different medical imaging modalities, showcasing its versatility.

Conclusions:

• The proposed DFQ framework effectively mitigates channel redundancy in deep representations for medical image analysis.
• This approach significantly improves the generalization ability of models across diverse medical imaging domains and scanner types.
• The findings suggest a promising direction for robust medical image representation learning in real-world clinical settings.