The Utilization Of Convolutional Neural Network On The Exploration Of NSCLC Biological Characteristics Through Chest CT Images

Background Non-small cell lung cancer(NSCLC)responds poorly to various treatment modalities and results in a lethal prognosis.The complicated biological characteristics and unpredictable trajectory of NSCLC make it unfeasible to individualize treatment for each patient and in further for each tumor lesion.Pathological and genetic examination of tissues and peripheral blood provides clinicians limited biological information about NSCLC.Nevertheless,pathological examination cannot be implemented repeatedly due to its invasiveness.Liquid biopsy of peripheral blood cannot describe the intra-and inter-heterogeneity of tumors for its imcompetence of considering anatomic information.Relatively high false-negative rate also makes it difficult to describe biological characteristics of NSCLC accurately.In contrast,convolutional neural network(CNN)on CT images could describe biological characteristics of NSCLC easily and conveniently and detect heterogeneity of tumors with account of anatomic information.Therefore,CNN could complement the shortcomings of pathological and genetic examination to some extent.Our study aims to utilize CNN to explore whether it could accurately describe the biological characteristics of NSCLC,including the detection of EGFR mutations and prediction of progression-free survival(PFS)of tyrosine kinase inhibitors(TKI).Materials and methods 1.Patients of stage IIIB&IV receiving first-line TKI treatment were retrospectively reviewed.Gender,smoking history,EGFR mutation site and metastasis patterns(oligometastasis and systemic metastasis)were analyzed to explore predictive factors of PFS and progression patterns(oligoprogression and systemic progression).2.Lung adenocarcinoma patients of stages I-IV were retrospectively collected and assigned randomly to training set and validation set according to patients' clinical characteristics.The CT images and EGFR gene status were utilized to train the CNN model and radiomic model.The two models were tested on validation set to ascertain their efficacy.3.A basic CNN model was built to distinguish malignancy of solid pulmonary nodules with a large training data size,which was then transferred to and fine-tuned by the data of training set of part two.Finally the fine-tuned CNN model was tested on the validation set of part two to see whether transfer learning could improve the ability of CNN to EGFR mutations.4.Chest CT images and PFS of patients receiving first-line TKI treatment were utilized to train a CNN model which is supposed to predict PFS for patients.Results 1.270 patients of stage IIIB&IV receiving first-line TKIs were collected to analyze predictive factors of PFS and 212 patients for progression pattern.The median PFS of total patients was 9 months(0.5-36 mos).PFS was correlated with EGFR mutation site(p=0.031)and the presence of cerebral metastasis(p=0.022,HR=1.353).Median PFS of patients with exon 18,19,20,21 mutations were 1 mos,10 mos,1 mos and 8 mos respectively.PFS of patients with cerebral metastasis were shorter than patients without cerebral metastasis(8 mos vs.10 mos,p=0.022).Progression pattern was related with metastasis pattern.Patients with oligometastasis were inclined to occur oligoprogression(p<0.0001,OR=11.32).Patients with bone metastasis were less likely to occur oligoprogression(p=0.008,OR=0.43).2.Totally 1010 patients(510 with EGFR mutations and 500 wildtype)of stage I-IV with lung adenocarcinoma were collected.810 patients were assigned to training set and 200 patients to validation set.The AUC value of CNN and radiomic model to detect EGFR mutations were 0.810 and 0.740 respectively.The specificity and sensitivity of CNN model was 0.753 and 0.813 at best decision point respectively.CNN model performed better than radiomic model(p=0.0225).3.The foundation CNN model could distinguish malignancy of solid pulmonary nodules with AUC value,specificity and sensitivity of 0.941,0.865 and 0.849 respectively.After transferring and fine-tuning,that model could detect EGFR mutations with AUC value of 0.863,which was higher than the results of part two(AUC=0.810).4.Totally 348 patients of stage IIIB&IV receiving first-line TKIs treatment were collected.Their treatment-naive chest CT images and PFS were utilized to build CNN model.Median PFS of 9 months divided total patients into short PFS group(?9 mos)and long PFS group(>9 mos).Total patients were allocated randomly into training set(278 patients)and validation set(70 patients).The AUC value,specificity and sensitivity of the established model to predict whether PFS was over 9 mos were 0.647,0.641 and 0.615 respectively.Conclusion 1.PFS was correlated with EGFR mutation site and cerebral metastasis.Progression pattern was correlated with metastasis pattern.2.CNN model could detect EGFR mutations with higher efficacy than radiomic model.3.The established CNN model could distinguish malignancy of solid lung nodules and improved the efficacy of CNN to detect EGFR mutations.4.The established CNN model could tentatively predict PFS of patients receiving first-line TKIs treatment.