Prognostic Prediction Of Locally Advanced Gastric Cancer Based On Deep Learning

Objective and PurposeThe prognosis of patients with gastric cancer is significantly different,so it is an urgent clinical problem to predict the prognosis of patients with gastric cancer early and guide the formulation of individualized treatment plan.The purpose of this study is to establish and verify the prognosis prediction model of 1-year progression-free survival rate（PFS）of LAGC patients based on CT deep learning characteristics,and to stratify the risk of tumor recurrence,so that patients can get individualized treatment plan at the initial treatment.MethodsThis study retrospectively collected the patients who underwent radical gastrectomy in the China-Japan Union Hospital of Jilin University,2017 to December,2020 and were pathologically confirmed as LAGC,but did not receive neoadjuvant chemotherapy before operation,and obtained preoperative clinical information and CT enhanced images.The baseline clinical information of the training set was analyzed by univariate and multivariate logistic regression,and the clinical risk factors significantly related to prognosis were screened.For preoperative CT images,the region of interest is delineated at the largest level and the farthest level of tumor in portal vein CT images,and deep learning features are extracted.LASSO analysis is used to select deep learning features and construct deep learning labels.Finally,three prognosis prediction models are constructed by logistic regression: preoperative clinical information model,deep learning feature model and comprehensive model.Visualize the comprehensive model with forest diagram and Nomo diagram.Discriminant degree,calibration degree and clinical decision curve were used to evaluate the predictive efficacy of the models,and Delong test was used to compare the differences between the models.ResultsA total of 422 patients were enrolled in this study and randomly divided into training set（n=295）and internal verification set（n=127）.For the preoperative clinical information of patients,sex,tumor location,carcinoembryonic antigen,and presence of signet ring cells are the clinical risk factors that are significantly related to the prognosis（P < 0.05）.Based on the pre-operative CT images of patients,nine deep learning features were finally selected to construct a deep learning label,which was significantly correlated with the 1-year progression-free survival rate（PFS）of LAGC patients（P < 0.001）.The comprehensive model constructed by incorporating deep learning labels and clinical information shows higher prognosis prediction performance than the clinical model.The area under the ROC curve（AUC）values of the comprehensive training set and the internal validation set are 0.771（95%CI,0.710～0.832）and 0.704（95%CI,0.583-0.826）,respectively,with the sensitivity of 65.8% and 66.7%,and the specificity of 74.3%.Both the training set and the verification set have good discrimination and calibration.The curve analysis shows that the prediction model has good clinical practicability.ConclusionThe characteristics of CT deep learning can predict the 1-year progression-free survival rate of patients with LAGC.The comprehensive prediction model combined with clinical information and deep learning labels can individually predict the prognosis of patients with LAGC after radical operation before operation,which is helpful to guide clinical individualized treatment strategies and improve the prognosis.