Training deep learning algorithms with weakly labeled pneumonia chest X-ray data for COVID-19 detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• The COVID-19 pandemic caused by SARS-CoV-2 has overwhelmed healthcare systems, increasing the demand for rapid diagnostic tools.
• Chest X-rays (CXRs) are crucial for assessing COVID-19 related pneumonia, but a lack of sufficient training data hinders the development of automated decision support systems.
• Radiology departments face significant burdens due to the high volume of CXR interpretations required.

Approach:

• This study addresses the data scarcity by utilizing weakly-labeled CXR images from public collections to augment training datasets.
• Convolutional neural network (CNN) algorithms were trained using a stage-wise approach on these augmented datasets.
• The goal was to develop and improve automated tools for detecting COVID-19 infections in CXRs.

Key Points:

• Weakly-labeled data augmentation significantly improved the performance of CNN models compared to non-augmented training.
• Augmentation expanded the feature space, enhancing inter-class discrimination and reducing generalization error.
• Training with COVID-19 specific CXRs demonstrated superior performance over non-COVID-19 viral pneumonia datasets.

Conclusions:

• Weakly-labeled data augmentation is an effective strategy to enhance the performance of AI models for COVID-19 detection in CXRs.
• COVID-19 associated viral pneumonia exhibits distinct patterns in CXRs compared to other pneumonias.
• The findings support the development of robust, AI-powered diagnostic aids to alleviate the strain on radiological services during pandemics.