A Study On The Identification And Prognostic Value Of Tertiary Lymphoid Structures In Lung Cancer Based On Pathomics Features

Objective: We developed a model for the identification of mature TLS in NSCLC by extracting the pathomic features of TLS and realised the accurate identification of mature TLS.The study preliminarily evaluated the role of TLS pathomic features on the prognostic value of NSCLC.Methods: 216 cases with postoperative NSCLC with positive TLS in the TCGA database and Meishan Cancer Hospital were enrolled,including 169 cases in the TCGA database as the training collection and 47 cases in Meishan Cancer Hospital as the validation collection.Whole Slide Imaging of postoperative NSCLC cases with positive TLS were collected in a uniform “.SVS” format.The regions of interest of the WSIs were outlined by Qupath software,and then the WSIs were cropped into patches of512*512 pixels with TLS in the region,with about 15-20 different patches per case.The RGB image format of each patch was normalized to a second-order grayscale image format using Python’s "opencv" package.The software Cellprofiler normalized the grayscale range of each patch and processed the patches with Gaussian filtering.Cell Profiler software extracted the TLS quantitative pathomic features of the processed patches.The Shapley value method was used to calculate the Shapley values of nonlinear features,and the Boruta algorithm calculated the average and standard deviation of the Shapley values for features selection.A Gaussian kernel function was used as the kernel function for the nonlinear support vector machine,and the "sklearn" package for Python was used to train and develop the mature TLS identification model.A Gaussian kernel function was used as the kernel function for the nonlinear support vector machine,and the "sklearn" package for Python was used to train and develop the mature TLS identification model.The ideal penalty parameter C and the kernel function parameter γ were determined by grid search and cross-validation,and the path-score was obtained by averaging with ensemble learning.The receiver operating characteristic curve,calibration curve,and decision curve were used to evaluate the efficacy of the mature TLS identification model.Cox proportional risk regression was used for univariate and multivariate analysis of NSCLC.The statistically significant variables in the univariate analysis were used to construct the NSCLC prognostic model,and Nomogram was plotted to facilitate the visualisation of the model.The Concordance Index,calibration curve,and decision curve were used to evaluate the efficacy of the prognostic model for NSCLC.Results: 1.The median age of the training and validation collections was 67 years(age range: 39-85 years),and the clinical stages were 111 cases of stage I,55 cases of stage II,42 cases of stage III,and 8 cases of stage IV.The TCGA database enrolled 169TLS-positive cases of NSCLC,including 68 cases of mature TLS and 101 cases of immature TLS.Meishan Cancer Hospital enrolled 47 cases,10 cases of mature TLS and37 cases of immature TLS.2.The Boruta algorithm calculated a maximum mean contribution value of2.16569125 for the Shapley value and selected seven pathomic features associated with mature TLS.3.Training and developing a mature TLS identification model based on a nonlinear support vector machine with a Gaussian kernel function.The ideal parameters of model were C=0.0156 and γ=0.024.The cut-off value of the path-score was determined by the Youden index to be 0.473.The ROC curves indicated that the built identification model for mature TLS had significant predictive efficacy with the area under the curve for the training collection of 0.934;and the area under the curve for the validation collection of 0.873.4.A prognostic model based on the pathomic features of TLS in lung cancer predicts long-term survival.The multidimensional Nomogram based on path-score,smoking,age,clinical stage,and N-stage built by the multidimensional Cox regression model had a concordance index of 0.672(0.639-0.705)for the training collection and0.730(0.668-0.793)for the validation collection.The probability of the 2-year,3-year,and 5-year survival of patients can be predicted visually.Conclusion:(1)The study found seven pathomic features closely associated with mature TLS.(2)The study developed an identification model of mature TLS in lung cancer based on pathomic features.(3)Path-score was an independent prognostic factor for NSCLC.A predictive Nomogram for survival associated with TLS pathomic features of NSCLC was developed,which was predictive of later survival of lung cancer patients.