Leveraging curriculum learning to address out-of-distribution data and inter-observer variability for lung nodule diagnostic interpretation

Reaching inter-observer consensus when diagnostically interpretating lung nodules is challenging for complex cases. When training, testing, and validating computer-aided diagnosis (CAD) systems, this diagnosis consensus is typically used as the class label to guide the learning. Furthermore, different imaging acquisition parameters used to acquire image data can cause data shifts, i.e. variations in the lung nodules’ appearances. These data shifts result in Out-of-Distribution (OoD) data, which are more challenging for algorithms to learn. In our study we aim to improve malignancy prediction for lung nodules using a novel curriculum learning (CL) approach, an incremental learning strategy, that starts to learn using In-Distribution (ID) cases and then adds iteratively easy to hard cases determined based on their OoD score. We show that the CL approach improves the classification of the malignant cases even for the out-of-distribution (OoD) samples and when the inter-observer interpretation variability is high. Our CL approach accurately classified 86.8% of malignant and 90.4% of benign high-variability nodules within specific OoD score thresholds (≤0.61 for malignant, ≤0.98 for benign). Cases outside these thresholds require data augmentation for improved learning.