Incremental learning method for lung nodule detection based on EWC and feature distillation

Abstract

The early manifestations of lung cancer are pulmonary nodules, and a computer aided diagnosis (CAD) model can help doctors to diagnose pulmonary nodules. However, when some new samples arrives, the existing model incremental update methods have problems such as catastrophic forgetting and high update spatiotemporal cost. To overcome the above problems, this paper proposes an incremental learning method for lung nodule detection based on elastic weight Consolidation (EWC) and feature distillation. First, lung parenchyma is segmented on CT images; secondly, the original model is optimized based on the elastic weight integration algorithm, the Fisher information matrix is simplified to a parameter weight matrix to reduce the amount of calculation, and then a regularization penalty term is introduced into the update process based on the parameter weight matrix The loss function limits the changes of important parameters to obtain the incrementally updated optimization model; finally, the original model is used as the teacher network, and the optimization model is used as the student network to calculate the feature distillation loss to improve the performance of the optimization model, which is realized without forgetting the learned knowledge. Sufficient fit to new samples. The method in this paper is cross-validated on the LUNA16 dataset. The results show that the method in this paper improves the incremental learning ability of the model, achieves good results in sensitivity and accuracy, and achieves low false positives and reduces space-time overhead.