Deep Active Learning for Lung Disease Severity Classification from Chest X-rays: Learning with Less Data in the...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Class imbalance and the need for extensive labeled data pose challenges in medical image analysis.
• Accurate classification of lung disease severity from chest X-rays (CXRs) is crucial for patient management.

Purpose of the Study:

• To reduce the amount of labeled data required for lung disease severity classification from CXRs.
• To address class imbalance in CXR datasets using deep active learning and Bayesian Neural Network (BNN) approximation.

Main Methods:

• Retrospective collection of 2319 CXRs from 963 COVID-19 patients.
• Application of deep active learning with BNN approximation and weighted loss function for disease severity classification.
• Evaluation using accuracy, AU-ROC, and AU-PRC, comparing various acquisition functions.

Main Results:

• Least Confidence strategy achieved 92.8% accuracy (AU-ROC, 0.95) in binary classification using only 9.24% of training data.
• Mean STD strategy achieved 70.5% accuracy (AU-ROC, 0.85) in multi-class classification using 21.87% of labeled data.
• Outperformed complex acquisition functions, significantly reducing labeling needs.

Conclusions:

• Deep active learning with BNN approximation and weighted loss effectively minimizes labeled data requirements for CXR analysis.
• This approach maintains or enhances diagnostic performance in lung disease severity classification, even with class imbalance.