Prediction Method Of EGFR Gene Mutation In Non-small Cell Lung Cancer Based On CT/PET Dual Modality

As a malignant tumor,lung cancer gradually corrodes human health and brings a huge threat to human survival.In our country,it ranks first in the number of cases and deaths every year,and the incidence of non-small cell lung cancer(NSCLC)has reached more than 80% of the incidence of lung cancer.Targeted drugs based on non-small cell lung cancer driver genes have become one of the most interesting treatments due to their reliable efficacy and mild side effects.Among the driver genes,the epidermal growth factor receptor(EGFR)gene is currently the target gene with the highest mutation rate and one of the most frequently used target genes for the treatment of non-small cell lung cancer.Only people who are sensitive to EGFR mutations can benefit from their targeted drugs.Therefore,EGFR mutation detection has become the primary requirement for clinical application of targeted drugs,especially the results of mutation detection.This thesis intends to non-invasively and automatically predict the mutation status of the patient’s EGFR gene from the patient’s computed tomography(CT)and positron emission computed tomography(PET)images by using a deep learning method.The research work done in this thesis mainly focuses on the following two points:Due to the insufficient number of studies on deep learning prediction methods for EGFR gene mutation status in non-small cell lung cancer at this stage,and all of them use a single modality of patient CT images as the input of the network.Therefore,this thesis proposes an improved Res Net for dual modality non-small cell lung cancer EGFR gene mutation prediction method.By improving the supervised Res Net-50 network,the algorithm uses the patient’s CT and PET dual modality image data as the input of the neural network on the non-small cell lung cancer EGFR gene mutation dataset,and fuses the patient’s clinical information features to help the network perform prediction of EGFR gene mutation status.Experiment’s results suggest that this method can predict the EGFR gene mutation status accurately,indicating its feasibility in helping clinical decision-making.Most of the current studies on the deep learning prediction of EGFR mutations in nonsmall-cell lung are based on supervised learning,which relies on a large number of expert hand labeled datasets.However,in practice,it will cost a lot of manpower and material resources for experts to manually label data,resulting in insufficient number of labeled data sets to support supervised learning in the medical and clinical field.In response to this problem,this thesis proposes a BYOL-based dual modality EGFR mutation status prediction method for non-small cell lung cancer.The algorithm modifies the self-supervised BYOL network,increases the number of layers of nonlinear multi-layer perceptrons in the network projection layer,and fuses patient CT and PET dual modality image data as the input of the network,then predict negative and positive cases without requiring large datasets of labeled patient medical images.Experimental results show that the improved BYOL network proposed in this thesis only needs a small number of annotated patient datasets to obtain more accurate detection results than some traditional supervised methods,demonstrating its potential to help clinical decision-making.