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

Reliable automated decisions are important in health diagnostics, especially for critical applications like disease diagnosis. Traditional machine learning models often produce uncertain predictions, which can have severe consequences in healthcare settings. To enhance diagnostic reliability, this paper introduces a Learning to Reject (L2R) method for COVID-19 classification from CT scans. L2R integrates data augmentation, calibrated confidence scores, and a rejection mechanism, enabling the model to abstain from uncertain predictions and reduce misclassifications. Our experimental results demonstrate that our L2R model achieves 100% accuracy with a low rejection rate, outperforming baseline models and significantly reducing diagnostic risk. This improved reliability underscores L2R's potential for crucial clinical applications and highlights its importance for future pandemic preparedness, where robust diagnostic tools are essential.