Decouple-Couple Network for Drug-Resistant EGFR Mutation Subtype Prediction with Lung Cancer CT Images

Abstract

Epidermal growth factor receptor (EGFR)-targeted therapy has revolutionized the treatment of EGFR-mutant lung cancer. However, a part of patients (nearly 10%) with mutated EGFR harbor drug-resistant mutation (DRM) subtypes. Although computed tomography images and deep learning have shown promising results in non-invasively predicting EGFR genotype, which may not be suitable to identify the DRM subtypes due to the imbalanced data distribution and the intra-class diversity of majority class. Hence, we propose a novel decouple-couple network (DCNet) to identify the DRM subtypes. Our DCNet firstly decouples the features of majority class as multiple prototypes, and then couple the prototypes of each class as one prototype for further classification. Meanwhile, the decouple-couple procedure is optimized jointly based on updated similarity score and prototypical contrastive learning. Furthermore, we collect a large CT dataset including 1232 EGFR-mutant lung cancer patients and the DCNet achieved sensitivity over 0.6, which improves largely than the state-of-the-art methods.