Prognostic Analysis And Prediction Of Acquired Drug Resistance Subtypes In Patients With Lung Adenocarcinoma Treated With EGFR-TKI Based On Artificial Intelligence

Lung cancer is the most common cancer in the world and the leading cause of cancer death.In recent years,although remarkable progress has been made in treatment,the survival rate of lung cancer is still very low in the world and even in China because most patients are diagnosed with advanced lung cancer and lung cancer has a high risk of recurrence and metastasis.At present,the treatment of lung cancer has entered the era of precision therapy,and the clinical application of EGFR tyrosine kinase inhibitor molecular targeting drugs has become a milestone event in the history of lung cancer treatment.The majority of EGFR mutation-positive lung adenocarcinoma patients benefit from this,extending their survival and improving their quality of life.However,the onset of acquired resistance interrupts this clinical benefit and leads to poor prognosis.At present,clinical monitoring of acquired drug resistance mainly relies on follow-up of computed tomography,but personalized follow-up strategies are not adopted according to individual differences of patients.This study is expected to use artificial intelligence to identify prognostication-related image features in CT images of patients’ internal and peritumoral tissues,analyze the information of heterogeneous changes before and after tumor treatment,predict disease progression and drug resistance subtypes in advance,and provide valuable information for the development of personalized clinical follow-up strategies.Part One Prognostic Analysis of Patients with Lung Adenocarcinoma Treated with EGFR-TKI Based on Serial-CT RadiomicsObjective To evaluate the value of time-serial CT radiomics features in predicting progression-free survival(PFS)and stratifying progressive risk for lung adenocarcinoma(LUAD)patients after epidermal growth factor receptortyrosine kinase inhibitors(EGFR-TKIs)therapy.Methods A total of 172 lung adenocarcinoma patients who were pathologically confirmed and treated with EGFR-TKI between January 1,2015 and September 30,2020 were enrolled.Among them,one hundred and thirty-one lung adenocarcinoma patients from our institute formed an internal training cohort and 41 lung adenocarcinoma patients from two independent medical institutes(24 patients from Institute 1 and 17 patients from Institute 2)formed an external validation cohort,respectively.The clinical data of 131 cases from the internal training cohort were retrospectively analyzed using one-way Cox regression analysis and clinical characteristics were selected,then clinical model was established.CT radiomics features were extracted from the primary tumor lesion and peri-tumor region from the CT images before treatment and the first followup after treatment and the most important radiomics features for predicting PFS were obtained by dimensionality reduction,and multi-time-point integrated(intratumor plus peri-tumor)radiomics signature,multi-time-point intratumor radiomics signature and baseline radiomics signature were established using Cox regression analysis,respectively.A combined model including clinical factors and multi-time-point integrated radiomics signature was developed.The aforementioned five models were applied to predict PFS for each patient separately and to obtain concordance index,which was validated in an external validation cohort.Paired nonparametric tests were used to compare the prediction accuracy between the models.The optimal cut-off value of the radiomics score was obtained by using the Kaplan-Meier log-rank test in survival analysis.The difference in survival between the two groups was observed by Kaplan-Meier curve.The receiver operating characteristic(ROC)curve and the area under curve(AUC)were used to assess the predictive efficacy of radiomics signature for predicting PFS at different time points.Results In the internal training cohort and the external validation cohort,the C-index was obtained by the multi-time-point integrated radiomics signature(Modelall)predicting PFS respectively(internal cohort: 0.78,95% confidence interval [0.72,0.84] and external cohort: 0.72,95% confidence interval [0.60,0.84].The C-index obtained by the clinical model(Modelclinical)for predicting PFS was 0.55 with 95% confidence interval [0.49,0.62](internal cohort)and 0.54 with 95% confidence interval [0.42,0.66](external cohort),respectively.The results showed that the predictive performance of Modelall was significantly better than Modelclinical(p<0.001).The C-index obtained for the combined model(Modelcombined)consisting of the clinical model combined with the integrated radiomics signature to predict PFS was 0.77 with 95% confidence interval [0.71,0.84](internal cohort)and 0.72 with 95% confidence interval [0.60,0.84](external cohort),respectively.The results suggested that clinical factors combined with radiomics signature did not improve the predictive performance of the model(internal: p= 0.713;external: p= 0.756).The C-index obtained by baseline radiomic signature(Modelbaseline)for predicting PFS was 0.66 with 95% confidence interval [0.60,0.73](internal cohort)and external: 0.57 with 95% confidence interval [0.50,0.69](external cohort),respectively.The C-index obtained by the intratumoral radiomics signature alone(Modelintratumor)to predict PFS was 0.71 with 95% confidence interval [0.65,0.78](internal cohort)and 0.65 with 95% confidence interval [0.53,0.77](external cohort),respectively.The results showed that in the internal training cohort,Modelall outperformed Modelbaseline and Modelintratumor in prediction and was statistically significant(p<0.001 or p= 0.006,respectively);in the external validation cohort,Modelall’s predictive performance was higher than Modelbaseline and Modelintratumor but not statistically significant(p= 0.086 or p= 0.198,respectively).The optimal cut-off value for the radiomics score obtained using the maximally selected log-rank test was 49.86.In the internal training cohort,88 individuals in the rapid progression group had a median progression-free survival of 8.3 months;43 individuals in the slow progression group had a median progression-free survivor of 13.4 months.In the external validation cohort,24 individuals in the rapid progression group had a median progression-free survival of 9.1 months;17 individuals in the slow progression group had a median progression-free survivor of 13.2 months.The results showed that patients in the rapid progression group had shorter progression-free survival than those in the slow progression group in both the internal training cohort and the external validation cohort(internal: p<0.001;external: p=0.022).The time-dependent ROC curves demonstrated that the proposed radiomics signature was able to accurately consistently effective in predicting the probability of disease progression at different time points.The AUCs at 6,9 and 12 months ranged from 0.89,0.90,0.94 in the training cohort and 0.76,0.80,0.96 in the validation cohort.Conclusions The comprehensive radiomics signature can accurately predict the progression-free survival time of lung adenocarcinoma patients treated with EGFR-TKI and the cut-off value of radiomics score can effectively differentiate lung adenocarcinoma patients treated with EGFR-TKI with different progression risks,which will help to make clinical decision and develop personalized follow-up strategies for this group.Part Two Prediction of Subtypes of Acquired Drug Resistance in Patients with Lung Adenocarcinoma Treated by EGFR-TKI Based on Deep LearningObjective To investigate the predictive value of deep learning convolutional neural networks for acquired drug-resistant subtypes in lung adenocarcinoma patients treated with epidermal growth factor receptortyrosine kinase inhibitor.Methods A total of 142 lung adenocarcinoma patients who were pathologically confirmed and treated with EGFR-TKI between January 1,2015 and September 30,2020 were collected.Among them,109 lung adenocarcinoma patients from our institute formed an internal training cohort and 33 lung adenocarcinoma patients from two independent medical institutes(22 patients from Institute 1 and 11 patients from Institute 2)formed an external validation cohort,respectively.Acquired resistance subtypes were classified separately by two senior oncologists according to clinical patterns of EGFR tyrosine kinase inhibitor failure in advanced non-small cell lung cancer for all progressive patients enrolled in the study.The Kappa test was used to compare the concordance of the results between two observers.The final classification results jointlynegotiated by the two oncologists were used as the classification results for the internal training set of this study.A convolutional neural network based on pre-treatment CT images and a convolutional-recurrent neural network based on pre-and post-treatment CT images for drug resistance subtype prediction model were constructed.The size of the dataset is expanded by image enhancement techniques,and transfer learning was applied to enhance the feature learning capability of the network models for small sample medical image data.The network was trained in an internal training cohort,and an external validation cohort was applied to validate the predictive performance of the model,and the performance of the model with different parameter combinations was evaluated by macro-averaging and micro-averaging ROC curves and AUC and F1 scores,respectively.The predictive efficacy of the optimal models was compared by Mc Nemar test.The network structure was implemented in Python using Keras with a Tensorflow backend(Python 2.7,Keras 2.0.8,Tensorflow 1.3.0).Results The internal training set had 18 cases of dramatic progression,72 cases of gradual progression,and 19 cases of local progression;the external validation set had 3 cases of dramatic progression,28 cases of gradual progression,and 2 cases of local progression.The interobserver coor of the results was high(kappa value of 0.676,95% confidence interval: 0.556,0.796).After image fusion and enhancement,the internal training set was expanded to 3379 samples,which were randomly divided into 3041 training samples and 338 test samples in the ratio of 9:1.The CR fusion neural network had 6758 image training images.The baseline neural network had3379 image training images.The macro-averaging AUC of the optimal CR fusion neural network prediction model was 0.78 and the micro-averaging AUC was 0.82;the area under the ROC curve for predicting the three subtypes was 0.73 for dramatic progression,0.75 for gradual progression,and 0.79 for local progression.The macro-averaging AUC of the optimal baseline neural network prediction model was 0.86 and the micro-averaging AUC was 0.93;the area under the ROC curve for predicting the three subtypes was 0.78 for dramatic progression,0.86 for gradual progression,and 0.84 for local progression,respectively.A Chi-squared value of 2 and a p-value of 0.18 for the Mc Nemar test concluded that there was no difference between the predictive performance of the baseline and CR fusion neural networks.Conclusions The baseline neural network had the capability of accurately predicting different subtypes of acquired resistance in EGFRTKI-treated lung adenocarcinoma patients,which would help clinical decision-making of treatment and personalized follow-up strategies.