An ensemble learning method based on ordinal regression for COVID-19 diagnosis from chest CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Public Health

Background:

• Coronavirus disease 2019 (COVID-19) poses a significant global health threat.
• Pulmonary X-ray computed tomography (CT) is crucial for COVID-19 management.
• Traditional CT image diagnosis is slow and error-prone, hindering rapid screening.

Purpose of the Study:

• To develop a deep learning (DL) method for accurate and rapid COVID-19 screening using CT images.
• To address the limitation of traditional DL methods that ignore disease progression order.
• To improve classification accuracy by learning both intraclass and interclass relationships among COVID-19 phases.

Main Methods:

• Proposed an ensemble learning method based on ordinal regression to leverage disease progression information.
• Utilized multi-binary, neuron stick-breaking (NSB), and soft labels (SL) techniques.
• Ensembled ordinal outputs using median selection with a modified ResNet-18 backbone.

Main Results:

• Achieved a 22% increase in accuracy compared to traditional methods in 2-fold cross-validation.
• Demonstrated improvements in precision, recall, and F1-score.
• The proposed method showed superior classification performance on 172 confirmed COVID-19 cases.

Conclusions:

• The ordinal regression ensemble learning method effectively classifies COVID-19 phases from CT images.
• This approach enhances diagnostic accuracy and efficiency for COVID-19 screening.
• Findings can aid in establishing guidelines for COVID-19 chest CT image classification.