A novel multiple instance learning framework for COVID-19 severity assessment via data augmentation and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate COVID-19 severity assessment is crucial during the pandemic.
• Chest CT scans are valuable for COVID-19 diagnosis but face challenges like weak annotation and insufficient data for automated analysis.
• Existing methods for automated severity assessment using CT images are limited by data scarcity and annotation quality.

Purpose of the Study:

• To develop and evaluate a novel method for fast and accurate COVID-19 severity assessment from chest CT images.
• To overcome limitations of weak annotation and insufficient data in automated CT-based COVID-19 analysis.
• To improve the performance of deep learning models in classifying COVID-19 severity.

Main Methods:

• A three-component method integrating deep multiple instance learning with instance-level attention.
• Bag-level data augmentation using high-confidence instances to generate virtual data.
• Self-supervised learning pretext task to enhance the training process.
• Systematic evaluation on a dataset of 229 COVID-19 cases (50 severe, 179 non-severe).

Main Results:

• Achieved an average accuracy of 95.8% in COVID-19 severity assessment.
• Demonstrated high diagnostic performance with 93.6% sensitivity and 96.4% specificity.
• Outperformed previous methods in automated COVID-19 severity classification using CT images.

Conclusions:

• The proposed method effectively addresses challenges in automated COVID-19 severity assessment from CT scans.
• The integration of multiple instance learning, data augmentation, and self-supervised learning significantly enhances diagnostic accuracy.
• This approach offers a promising tool for rapid and reliable evaluation of COVID-19 severity, aiding clinical decision-making.