Improving domain generalization performance for medical image segmentation via random feature augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep convolutional neural networks (DCNNs) excel in medical image segmentation.
• Distribution discrepancies in medical images challenge DCNN robustness on unseen data.
• Existing domain generalization methods often require prior knowledge, limiting data diversity.

Purpose of the Study:

• To develop a novel method for enhancing DCNN generalization in medical image segmentation.
• To address the limitations of prior knowledge-dependent feature augmentation techniques.
• To improve the robustness of DCNNs against variations in medical imaging data.

Main Methods:

• Proposed a random feature augmentation (RFA) method to diversify source domain data at the feature level without prior knowledge.
• RFA perturbs domain-specific information while preserving domain-invariant information.
• Introduced a dual-branches invariant synergistic learning strategy to capture domain-invariant features from RFA-augmented data.

Main Results:

• Demonstrated superior performance of the proposed method over state-of-the-art domain generalization techniques.
• Evaluated on optic cup/disc segmentation (fundus images) and prostate segmentation (MRI images).
• The method effectively diversifies source domain data and enables learning of generalized representations.

Conclusions:

• The proposed RFA method and synergistic learning strategy significantly enhance DCNN generalization for medical image segmentation.
• This approach offers a robust solution for applying DCNNs to diverse clinical datasets.
• The method shows promise for improving diagnostic accuracy in various medical imaging applications.