Addressing cross-population domain shift in chest X-ray classification through supervised adversarial domain...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Artificial intelligence (AI) is vital for medical image analysis in healthcare.
• Machine learning models struggle with domain shift, limiting generalization across diverse patient populations.
• Chest X-ray classification faces challenges due to cross-population variations, particularly in underrepresented groups.

Purpose of the Study:

• To investigate domain shift issues in chest X-ray classification across different populations.
• To propose and evaluate a supervised adversarial domain adaptation (ADA) technique to address cross-population domain shift.
• To improve the performance of AI models on underrepresented datasets.

Main Methods:

• Analyzed domain shift impact using three source population datasets and a Nigerian chest X-ray dataset as the target.
• Developed a supervised adversarial domain adaptation (ADA) method involving a feature extractor and an adversarial domain discriminator.
• Trained the feature extractor on source domains and used adversarial training to create domain-invariant features.

Main Results:

• Significant performance discrepancies were observed when models trained on source domains were applied to the target Nigerian dataset.
• The proposed ADA technique demonstrated substantial improvements in chest X-ray classification on the Nigerian dataset.
• The model achieved 90.08% accuracy and 96% AUC, outperforming multi-task learning (MTL) and continual learning (CL).

Conclusions:

• Domain shift poses a significant challenge for AI in medical imaging, especially across diverse populations.
• Supervised adversarial domain adaptation (ADA) effectively creates domain-invariant features, mitigating cross-population disparities.
• Developing domain-aware AI models is crucial for equitable and effective healthcare diagnostics.