Correcting data imbalance for semi-supervised COVID-19 detection using X-ray chest images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Early identification of coronavirus disease (COVID-19) carriers is crucial for disease control.
• Deep learning on chest X-rays offers a potential pre-diagnostic tool, but requires large, balanced datasets.
• New outbreaks present challenges with small, imbalanced datasets, hindering deep learning model performance.

Purpose of the Study:

• To evaluate the MixMatch semi-supervised deep learning architecture with limited and imbalanced chest X-ray datasets for COVID-19 detection.
• To address the challenge of data imbalance in deep learning for novel viral diseases.
• To propose and validate a simple data imbalance correction method for improved classification accuracy.

Main Methods:

• Utilized the MixMatch semi-supervised learning architecture.
• Evaluated model performance on highly imbalanced datasets with very few labeled observations.
• Proposed a re-weighting strategy within the loss function to correct for data imbalance, assigning higher weights to under-represented classes.
• Used pseudo and augmented labels for unlabeled data to determine appropriate weights.

Main Results:

• Demonstrated the significant negative impact of data imbalance on deep learning model accuracy for COVID-19 detection.
• The proposed re-weighting method improved classification accuracy by up to 18% compared to the standard MixMatch algorithm.
• Successfully tested the approach on binary (COVID-19 vs. normal) and multi-class (COVID-19, pneumonia, normal) datasets, including a new dataset from Costa Rican patients.

Conclusions:

• Semi-supervised learning with data imbalance correction is effective for COVID-19 detection using chest X-rays, even with limited data.
• The proposed re-weighting technique offers a simple yet powerful solution to improve deep learning model performance in resource-constrained scenarios.
• This approach holds promise for rapid pre-diagnostic tool development during emerging viral outbreaks.