Discriminating Serosal Invasion And Its Transfer Application In Gastric Cancer Based On Deep Learning

Objective: Serosal invasion of gastric cancer increases the risk of peritoneal dissemination and leads to poor prognosis.Current clinical practice guidelines recommend that patients with serosal invasion receive neoadjuvant chemotherapy before surgery,so the assessment of serosal invasion before surgery will determine different treatment modalities.However,the accuracy of imaging methods such as Computed Tomography(CT)and Endoscopic Ultrasonography(EUS)on T staging and serosal invasion before surgery does not meet the clinical needs.Moreover,for the patients with peritoneal metastases,traditional preoperative imaging methods,such as CT,have low sensitivity and cannot detect occult peritoneal metastases.For patients with serosal invasion,the incidence of occult peritoneal metastases was more than four times higher than patients without serosal invasion.Previous studies have developed the model with preoperative CT to identify occult peritoneal metastasis found by laparoscopic exploration and achieve high accuracy,but there is still a lack of research on identifying patients with positive peritoneal washing/ascites cytology.If such patients can be found before or during surgery,and treated with appropriate treatment methods,such as intraperitoneal chemotherapy and peritoneal lavage,can increase the survival rate and decreases the peritoneal recurrence rate.In addition,the macroscopic serosal appearance can be used as a predictor of peritoneal metastasis.However,this method is mainly dependent on the subjective observation by surgeons,and the judgment is easily affected by other factors.Therefore,building a model integrating preoperative CT and intraoperative serosal photographs may improve the predictive accuracy of positive peritoneal washing/ascites cytology.In clinical practice treatment,patients with serosal invasion are likely to benefit from neoadjuvant therapy than other patients.The high-accuracy model for predicting serosal invasion has laid the foundation to accurately select patients respond to neoadjuvant chemotherapy.However,not all patients with serosal invasion benefit from neoadjuvant chemotherapy.We still need to further search and predict the efficacy of neoadjuvant chemotherapy.Deep learning have made major breakthroughs in computer vision and are widely used in medical image processing.An emerging deep learning method,Vision Transformer(Vi T),extends the traditional Transformer in the field of Natural Language Processing(NLP)to image classification tasks.Vi T has outperformed Convolutional Neural Networks(CNN)in some situations.However,the effectiveness of Vi T in medical images still needs to be explored.Transfer learning,which improves the learning of new tasks by transferring knowledge from already learned related tasks,has become a popular optimization algorithm in deep learning.In addition,the use of Generative Adversarial Network(GAN)can augment the training set and solve the problem of insufficient number of cases in the training set.Methods: In this study,deep learning was first performed to preoperatively discriminate serosal invasion of gastric cancer with CT.2275 patients in The First Hospital of China Medical University(2039 in training set,236 in internal test set)and 696 patients in Liaoning Cancer Hospital were included.Based on the region of interest(ROI)of CT lesions manually annotated by researchers and radiologists,discriminative effects of Dense Net-169,Inception-v3,Inception-Resnet-v2,Res Net-50 and Vi T-B/32 on serosal invasion were compared.Subsequently,based on the CT images of 1928 patients(1720 in the training set and 208 in the internal test set)and 560 patients in the external test set,and the preoperative blood indicators with significant difference between serosal invasion and non-invasion patients,Dense Net-169+C,Inception-v3+C,Inception-Resnet-v2+C,Res Net-50+C and Vi T-B/32+C were constructed.Combining the best performing CNN model(Inception-v3+C)and Vi T model(Vi T-B/32+C),Ensemble Model+C was constructed,and the preoperative CT images and blood indicators were combined to discriminate serosal invasion and the performance of the models were evaluated in the internal and external test sets.In addition,a reproducibility experiment was carried out.Another oncologist independently annotated the tumor region on CT in the external test set,and the annotated ROI was input into the model to evaluate the performance of the model and explore whether the effect of the model was subject to inter-observer bias.Survival analysis was then performed to determine the prognostic evaluation performance of the model.Second,transfer learning was applied to transfer the serosal invasion model to the peritoneal washings/ascites cytology prediction model.Based on the CT images and preoperative blood indicators of 620 patients in The First Hospital of China Medical University(496 in the training set and 124 in the test set),the best performing model in the first part(Ensemble Model+C)was transferred to predict the results of peritoneal washing/ascites cytology through transfer learning.To further improve the accuracy,this study collected intraoperative serosal photographs of patients,and compared the effects of Dense Net-169,Inception-v3,Inception Resnet-v2,Res Net-50 and Vi T-B/32 on serosal invasion,the model was transferred to predict the results of peritoneal washing/ascites cytology,and the model with the best effect was selected and combined with the outputs of the preoperative Ensemble Model+C to predict the results of peritoneal washing/ascites cytology during surgery.Finally,the response of neoadjuvant chemotherapy for gastric cancer was predicted based on the primary tumor and lymph node metastases in CT.The CT images were included from 188 patients in Liaoning Cancer Hospital and 94 patients in The First Hospital of China Medical University before neoadjuvant chemotherapy.Based on the CT images of primary tumor and lymph node metastases,Deep Convolutional Generative Adversarial Network(DCGAN)performs data argumentation in a 1:3 ratio.Deep learning models for the prediction of neoadjuvant chemotherapy efficacy was constructed on Dense Net-169,Inception-v3,Inception Resnet-v2,Res Net-50 and Vi T-B/32 respectively,with the images generated by DCGAN,and the models were transferred to the CT images of the primary tumor and lymph node metastases,respectively.The accuracy and AUC value of the model were calculated by combining the outputs of the primary tumor and lymph node metastasis models.With the help of Ensemble Model+C,the subgroup of serosal invasion was identified,and this study evaluated predictive effect of the model on chemotherapy response in patients with serosal invasion.Finally,the study used Cox regression to analyze the prognostic evaluation performance of the model.Results: 1.Preoperative discrimination of serosal invasion in gastric cancer This part firstly builds a deep learning model for serosal invasion discrimination based on CT images.Dense Net-169 has the highest accuracy in the internal test set,with an accuracy rate of 89.4%,and Inception-v3 has the highest accuracy in the external test set,with an accuracy rate of 87.5%.After combining CT with blood indicators related to serosal invasion,the Ensemble Model+C constructed by combining Inception-v3+C and Vi TB/32+C performed best,with an accuracy of 91.8 % and the Area Under Curve(AUC)of the Receiver Operating Characteristic Curve(ROC)of 0.974(95% CI: 0.955-0.992)in the internal test set,the accuracy of 90.9% and the AUC of 0.950(95%CI: 0.931-0.969)in the external test set.The reproducibility experiment shows that the accuracy of Ensemble Model+C is 87.5%.The Kaplan-Meier curve log-rank test indicated that Ensemble Model+C could predict postoperative survival(P<0.001).The Concordance Index(Cindex)was 0.713(95% CI: 0.686-0.739)in the training set and 0.754(95%CI:0.709-0.798)in the external test set.2.Predicting the results of peritoneal washing/ascites cytology Ensemble Model+C in serosal invasion discrimination task was transferred to predict the results of peritoneal washing/ascites cytology,a total of 108 patients in test set with preoperative CT,blood indicators and intraoperative serosal photographs were included.the accuracy rate reached 89.8%,and the AUC was 0.952(95%CI : 0.907-0.997).Intraoperative gastric serosal photographs can predict serosal invasion and transferred to predict the results of peritoneal washing/ascites cytology.Dense Net-169 had the highest accuracy(83.3%).When the two outputs were combined,the accuracy of the model reached 91.7%,and the AUC reached 0.965(95%CI:0.934-0.995).3.Prediction of the response of neoadjuvant chemotherapy for gastric cancer Based on the primary tumor CT images of gastric cancer to predict the response of neoadjuvant chemotherapy,Res Net-50 achieved the highest accuracy rate of 81.9% in the test set.When the primary tumor and lymph node metastasis outputs were combined,there is an improvement in the AUC value,and Inception-Res Net-v2 achieved the highest AUC value of 0.918(95% CI: 0.859-0.978).When we only included patients with serosal invasion and only considered the primary tumor,Res Net-50 achieved the highest accuracy rate of 81.2%.When the model combining the outputs of the primary tumor and the lymph node metastasis,the AUC for Inception-Res Net-v2 is 0.902(95%CI:0.832-0.972).Conclusions: 1.Ensemble Model+C has a good performance in preoperative discrimination of serosal invasion in gastric cancer based on CT images and blood test indicators,and the model has a predictive effect on prognosis.2.The Ensemble Model+C can adopt the transfer learning and combined with the deep learning model of intraoperative serosal photographs to predict the results of peritoneal washing/ascites cytology before or during surgery.3.Deep learning models can be constructed based on the primary tumor and lymph node metastases of gastric cancer,respectively,and the two model outputs can be combined to predict the response of neoadjuvant chemotherapy.