Guideline-Based Additive Explanation for Computer-Aided Diagnosis of Lung Nodules

摘要

Machine Learning (ML) models have achieved remarkable predictive capability in Computer-Aided Diagnosis (CAD) systems. However, a problem of such models is that they are regarded as black-box models and lack of an explicit representation. In this work, a Guideline-based Additive explanation (GAX) framework is proposed for interpreting ML-based CAD systems. A medical guideline standardizes decision making in disease diagnosis. The idea of GAX is generating understandable explanations according to the criteria of the guideline. It contains two steps: anatomical features defined on the basis of the guideline are first generated using rule-based segmentation and anatomical regularities, and perturbation-based analysis is then used for calculating the importance of each feature. In addition, global explanation is also obtained by analyzing the entire dataset, where measurements are calculated from anatomical features, and a figure containing the overview of which measurements are important is generated. The proposed GAX is evaluated on a lung CT image dataset. The results demonstrate that GAX can provide understandable explanations to gain trust in clinical practice, and also present data bias for users to further improve the model.