Learning to Generalize Towards Unseen Domains via a Content-Aware Style Invariant Model for Disease Detection From...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Distribution discrepancy poses a challenge for intelligent imaging, causing CNNs to focus on image style over content.
• Human radiologists effectively generalize across chest X-ray (CXR) domains by learning visual cues.

Purpose of the Study:

• To develop a robust method for cross-domain generalization in chest X-ray analysis.
• To enhance the performance of deep learning models on unseen CXR datasets by addressing style bias.

Main Methods:

• Employed on-the-fly style randomization modules at image (SRM-IL) and feature (SRM-FL) levels.
• SRM-IL samples style statistics from the possible value range of CXR images for diverse augmentations.
• SRM-FL utilizes pixel-wise learnable parameters for style embeddings and consistency regularizations are applied to global semantic features and predictive distributions.

Main Results:

• The proposed method achieved superior AUCs on unseen CXR datasets (BRAX, VinDr-CXR, NIH chest X-ray14) compared to state-of-the-art models.
• Statistically significant improvements were observed in thoracic disease classification.
• Achieved 77.32±0.35, 88.38±0.19, 82.63±0.13 AUCs(%) on unseen domains.

Conclusions:

• On-the-fly style randomization effectively creates style-perturbed features while preserving content integrity.
• The method enhances model robustness and generalization capabilities for cross-domain CXR analysis.
• This approach offers a promising solution for improving AI diagnostic performance in medical imaging.