Preoperative CT-based Deep Learning Model For Predicting Lymphovascular Invasion In Clinical Stage IA Lung Adenocarcinoma

Purpose:To evaluate the predictive value of deep learning model of Clinical model,Transformer model and Clinical-Transformer model for the presence of lymphovascular invasion(LVI)of patients with stage IA lung adenocarcinoma before surgery,providing an important reference value for clinical decision-making and individualized treatment.Materials and methods:This study retrospectively analyzed the clinical,pathological and imaging data of patients diagnosed with lung adenocarcinoma after surgery in the First Affiliated Hospital of Guangzhou Medical University from January 1,2016 to July 30,2019,and in the Cancer Prevention Center of Sun Yat-sen University from January 1,2010 to December 31,2021.The data of the two centers were mixed and randomly divided into the training cohorts and the test cohorts according to the ratio of 8:2.The independent risk factors of lung adenocarcinoma LVI were screened by one-way and multiple logistic regression analysis.Trans UNet segmentation algorithm was used to segment the region of interests(ROIs),and then the image blocks were input to Res Net18 for pre-training.Then Vision Transformer architecture was used to extract the features.Finally,the selected clinical,CT imaging features and CT image deep learning features were input into the full connect neural network classifier to construct the models.The model was selected and evaluated by the method of five-fold cross-validation.The "report ROC" package in R software was used to draw the receiver operating characteristic curve(ROC)of different models.Area under curve(AUC),accuracy,specificity,sensitivity,positive predictive value(PPV),negative predictive value(NPV)and F1 score were calculated to evaluate the predictive efficacy of each deep learning model.De Long method was used to test and compare the statistical differences of AUC among different models.Results:A total of 1907 patients with clinical stage IA lung adenocarcinoma were included in this study,including 290 LVI-positive patients,1617 LVI-negative patients,with an average age of 58.4±9.9 years.They were divided into 1525 cases(232 LVI positive cases,1293 LVI negative cases)in the training cohorts and 382 cases(58 LVI-positive patients,324 LVI-negative patients)in the test cohorts.The results of multiple Logistics regression analysis showed that gender,smoking history,serum carcinoembryonic antigen level,pathological grade of lung adenocarcinoma,lymph node metastasis,visceral pleural invasion,nodule composition and peritumoral interstitial thickening were independent risk factors for clinical stage IA lung adenocarcinoma LVI status.And the risk of LVI increased with the increasing of lung adenocarcinoma pathological grade.In the study,there were three deep learning models constructed: Clinical model,Transformer model and Clinical-Transformer combined model.In both the training cohorts and test cohorts,the combined Clinically-Transformer model outperformed in the predictive efficacy for preoperative LVI status of lung adenocarcinoma,with AUCs of 0.862(95%CI: 0.839-0.885)and0.742(95%CI: 0.716-0.768),respectively.Transformer model ranked secondly,with the AUC were 0.834(95%CI: 0.809-0.859)and 0.728(95%CI: 0.702-0.754),respectively.The clinical model had the worst predictive power,and its training cohorts and test cohorts AUC were 0.831(95%CI: 0.807-0.856)and 0.713(95%CI:0.868-0.740),respectively.Conclusions:1.The study develops a preoperative deep learning prediction model to predict the presence of LVI in clinical stage IA lung adenocarcinoma.By combining clinical, pathological and CT imaging features,the Clinical-Transformer model achieves the highest prediction efficiency.In the future,the predictive efficacy of thismodel is expected to be further improved,making it a potential auxiliary diagnostic tool for noninvasive preoperative prediction of LVI status,and providing important reference value for clinical decision-making and individualized treatment.2.The incidence of LVI in clinical stage IA lung adenocarcinoma is closely related to the pathological grade of adenocarcinoma.With the higher pathological grade, the incidence of LVI increases,with the highest risk of micropapillary predominant adenocarcinoma and solid predominant adenocarcinoma.