Enhancing Reliability in COVID-19 Classification from CT scans using Learning to Reject.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Diagnostic Reliability

Background:

• Automated decisions in health diagnostics are critical, but traditional machine learning models often yield uncertain predictions.
• Uncertainty in healthcare can lead to severe consequences, necessitating more reliable diagnostic tools.

Purpose of the Study:

• To introduce a Learning to Reject (L2R) method for enhancing the reliability of COVID-19 classification from CT scans.
• To enable automated systems to abstain from uncertain predictions, thereby reducing misclassifications and diagnostic risk.

Main Methods:

• The Learning to Reject (L2R) method integrates data augmentation, calibrated confidence scores, and a rejection mechanism.
• This approach allows the model to identify and reject uncertain predictions, focusing on high-confidence classifications.

Main Results:

• The L2R model achieved 100% accuracy in COVID-19 classification from CT scans.
• The method demonstrated a low rejection rate while significantly outperforming baseline models.
• Diagnostic risk was substantially reduced due to the enhanced reliability of the model's predictions.

Conclusions:

• The L2R method significantly improves the reliability of automated diagnostic decisions in critical healthcare applications.
• This approach holds substantial potential for clinical use and is vital for future pandemic preparedness.
• Robust diagnostic tools like L2R are essential for accurate and safe disease classification.