Improved BYOL method for predicting epidermal growth factor receptor gene mutations in non-small cell lung cancer

The epidermal growth factor receptor (EGFR) mutation status can predict whether a patient has non-small cell lung cancer (NSCLC). A self-supervised EGFR gene mutation prediction method based on contrastive learning is proposed, which can distinguish between negative and positive images of the patient's lesion area input to the network, without requiring a large number of expert hand-labeled patient datasets. The self-supervised BYOL network was modified to increase the number of layers of the non-linear multilayer perceptron (MLP) of the network projection layer, and image data of the patient's CT and PET modalities were merged as the input of the network. Negative and positive medical records can be predicted without the need to annotate a large number of patient datasets. Using the non-small cell lung cancer EGFR gene mutation datasets, it is compared with traditional radiomics, supervised VGG-16 network, supervised ResNet-50 network, supervised Inception v3 network, and unsupervised transfer learning CAE. The experimental results show that the instance features of patient lesion area images learned from CT and PET images of patients using contrastive learning can be used to distinguish negative and positive cases, with an area under the curve (AUC) of 77%. The classification results improved by AUC of 7% compared to the traditional radiomics method, and by AUC of 5% compared to the classification results of the supervised VGG-16 network. The AUC is only 9% lower than that of supervised ResNet-50, without requiring a large number of expert hand-annotated datasets and large patient clinical datasets. The improved BYOL network proposed in this paper requires only a small number of labeled patient datasets to obtain more accurate prediction results than some traditional supervised methods, demonstrating its potential to help clinical decision-making. © 2022, Science Press. All right reserved.