Computer-aided Assessment Of Response To EGFR-KTI Therapy Based On MRI Of Brain Metastasis From Lung Cancer

Objective: Epidermal growth factor receptor(EGFR)-tyrosine kinase inhibitors(TKIs)therapy has been recognized as one of the most effective treatment strategies for lung cancer,and preoperative prediction of EGFR mutation status and response to EGFR-TKI in NSCLC patients with brain metastases(BM)is critical for clinical decision.This study investigates the value of MRI images of multi-modal brain metastases to predict EGFR mutation status and the response to EGFR-TKI based on radiomics and deep learning.Methods: A total of 310 patients with BM from lung cancer in two hospitals were included in this study.All patients underwent contrast-enhanced T1-weighted(T1C)and T2-weighted(T2W)MRI scans prior to treatment.This study includes three parts,the first part focuses on investigating the predictive value of whole tumor,the second part explores the predictive value of the brain-to-tumor interface(BIT),and the third part use the deep learning method to build an automated prediction system.The detailed methods were as follows: In the first part,radiomics features were extracted from the tumor active area(TAA)and peritumoral oedema area(POA)of each patient.The most predictive features were selected using the least absolute shrinkage and selection operator(LASSO).Then,a logistic regression algorithm was used to construct the radiomics signatures(RS).In the second part,the features were extracted from the 4-mm BTI and the BM,respectively.Feature selection was performed using the Mann-Whitney U test,LASSO and Akaike Information Criteria(AIC).Fusion models were constructed by combining the volume of peritumoral edema(VPE)and selected radiomics features through logistic regression.In the third part,an automated artificial intelligence EGFR system(EGFR-AIS)was proposed.The EGFR-AIS consists of two networks,the segmentation network with FC-Densenet backbone,using Leak Relu nonlinear activation function instead of Relu nonlinear activation function,and integrating an external attention mechanism to improve the accuracy.The classification network uses Densenet-121 as the backbone,and the fully connected layer was replaced by global average pooling(GAP)to reduce the number of parameters.Finally,the area under receiver operating characteristic(ROC)curves(AUC)are used to evaluate the model performance.Results: The first part of the results showed that for predicting EGFR mutation status,the constructed POA-and TAA-based RSs showed similar performance.The multi-region fusion model constructed by fusing the POA-and TAA-based RSs achieved the highest prediction performance with AUC of 0.896,0.856,and 0.889 in the training,internal validation,and external validation set,respectively.For predicting the response to EGFRTKI,the multi-region fusion model generated the highest prediction performance in the training(AUC=0.817),internal validation(AUC= 0.788),and external validation set(AUC=0.808),respectively.The second part of the results showed that VPE was helpful to predict the EGFR mutation status and the response to EGFR-TKI.The combined model constructed by fusing VPE with the radiomics features yielded the highest AUC values in the training(AUC=0.890),the internal validation(AUC=0.833),and the external validation(AUC=0.814)set for predicting EGFR mutation status.In predicting the response to EGFR-TKI,the combined model also achieved optimal prediction performance in the training(AUC=0.766),the internal validation(AUC=0.712),and the external validation(AUC=0.730)set,respectively.The third part of the results showed that the proposed EGFR-AIS system achieved excellent performance in predicting EGFR mutations with AUC values of 0.825 and 0.821 in the internal validation and external validation set,respectively.For predicting the response to EGFR-TKI,the AUC values of the proposed model were 0.723 in the internal validation and 0.711 in the external validation set.Conclusion: This study demonstrates the value of a multi-regional combined RSs based on BM for predicting the EGFR mutation status and response to EGFR-TKI.The combined model based on BTI-based RS with VPE was also valuable for predicting EGFR mutation status and response to EGFR-TKI.Unlike labor-intensive radiomics-based imaging models,the constructed MRI-based EGFR-AIS can provide a potentially non-invasive automated tool for preoperative prediction of EGFR mutations and response to EGFR-TKI.