Curriculum gDRO: Improving Lung Malignancy Classification through Robust Curriculum Task Learning

Deep learning models used in Computer-Aided Diagnosis (CAD) systems are often trained with Empirical Risk Minimization (ERM) loss. These models often achieve high overall classification accuracy but with lower classification accuracy on certain subgroups. In the context of lung nodule malignancy classification task, these atypical subgroups exist due to the lung cancer heterogeneity. In this study, we characterize lung nodule malignancy subgroups using the malignancy likelihood ratings given by radiologists and improve the worst subgroup performance by utilizing group Distributionally Robust Optimization (gDRO). However, we noticed that gDRO improves on worst subgroup performance from the benign category, which has less clinical importance than improving classification accuracy for a malignant subgroup. Therefore, we propose a novel curriculum gDRO training scheme that trains for an "easy" task (nodule malignancy is determinate or indeterminate for radiologists) first, then for a "hard" task (malignant, benign, or indeterminate nodule). Our results indicate that our approach boosts the worst group subclass accuracy from the malignant category, by up to 6 percentage points compared to standard methods that address and improve worst group classification performance.