Augmentation strategies for an imbalanced learning problem on a novel COVID-19 severity dataset

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Machine learning models for COVID-19 detection from chest X-rays are common.
• Binary classification models have limited treatment implications.
• Predicting COVID-19 severity is crucial for tailored medical interventions.

Purpose of the Study:

• To create and release one of the largest publicly available COVID-19 severity datasets.
• To establish a benchmark for deep learning-based COVID-19 severity classification.
• To investigate augmentation strategies for imbalanced severity data.

Main Methods:

• Compiled a dataset of 2358 COVID-19 positive chest X-ray images with severity scores (COVIDx8B dataset).
• Trained and evaluated deep learning models for severity classification.
• Implemented and tested data augmentation techniques targeting majority and minority classes.

Main Results:

• The newly created dataset serves as a benchmark for COVID-19 severity classification.
• Augmentation strategies significantly improved precision and recall for severe COVID-19 cases.
• Models demonstrated enhanced performance on rare, severe cases despite initial class imbalance.

Conclusions:

• The developed dataset and models offer a valuable starting point for further research.
• Augmentation techniques are effective in addressing class imbalance for severe disease prediction.
• Future work can optimize models for clinical application in resource allocation and treatment planning.