Viral Pneumonia Screening on Chest X-Rays Using Confidence-Aware Anomaly Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Epidemiology

Background:

• Viral pneumonia outbreaks necessitate rapid detection for epidemic control.
• Chest X-rays are valuable for screening when advanced imaging is unavailable.
• Novel viral mutations challenge existing classification models due to dataset shifts.

Purpose of the Study:

• To develop a robust method for differentiating viral pneumonia from non-viral pneumonia and healthy cases.
• To address the limitations of classification models facing novel viral mutations.
• To propose a one-class classification-based anomaly detection approach for viral pneumonia detection.

Main Methods:

• Formulated viral pneumonia detection as a one-class classification anomaly detection problem.
• Proposed the confidence-aware anomaly detection (CAAD) model with a shared feature extractor, anomaly detection, and confidence prediction modules.
• Treated all viral pneumonia cases as anomalies to enhance the one-class model's performance.

Main Results:

• The CAAD model outperformed binary classification models on the X-VIRAL dataset (5,977 viral pneumonia cases).
• On the X-COVID dataset, CAAD achieved an AUC of 83.61% and 71.70% sensitivity without fine-tuning.
• Performance on the X-COVID dataset was comparable to that of human radiologists.

Conclusions:

• The CAAD model offers an effective approach for viral pneumonia detection, particularly in the context of novel viral strains.
• Anomaly detection provides an advantage over binary classification by not requiring explicit modeling of individual viral pneumonia types.
• The model demonstrates potential for large-scale screening and epidemic prevention, even with limited resources.