Abstract
We introduce a new end-to-end computer aided detection and diagnosis systemfor lung cancer screening using low-dose CT scans. Our system is based on 3Dconvolutional neural networks and achieves state-of-the-art performance forboth lung nodule detection and malignancy classification tasks on the publiclyavailable LUNA16 and Kaggle Data Science Bowl challenges. Furthermore, wecharacterize model uncertainty in our system and show that we can use this toprovide well-calibrated classification probabilities for nodule detection andpatient malignancy diagnosis. To the best of our knowledge, model uncertaintyhas not been considered in the context of lung CT analysis before. Thesecalibrated probabilities informed by model uncertainty can be used forsubsequent risk-based decision making towards diagnostic interventions ordisease treatments, as we demonstrate using a probability-based patientreferral strategy to further improve our results.