Establishment Of An Artificial Intelligence-based Recognition Model For Mixed-type Gastric Cancer Pathological Images And Prediction Of Lymph Node Metastasis

ObjectiveEstablishing a recognition model for mixed-type gastric cancer based on pathological images using deep learning,quantifying the specific ratio of glandular differentiation component to undifferentiated component in advanced mixed-type gastric cancer,and assessing the correlation between this ratio and the risk of lymph node metastasis.Methods1.Establishing a recognition and automatic segmentation model for glandular differentiation components in advanced mixed-type gastric cancer using the U-Net architecture.2.Using the open-source pathology software QuPath to create a threshold model,quantifying the low-differentiation components in advanced mixed-type gastric cancer and segmenting them from normal tissue.3.Validating the two models above and quantifying the ratio of differentiation components to undifferentiated components in advanced mixed-type gastric cancer based on the output of the models,and exploring the relationship between this ratio and lymph node metastasis.ResultsAfter labeling the ROI region with 186 SVS format images,a total of 13003 ROI regions were obtained.U-Net model was used to train the identification and segmentation of adenoid differentiation components in advanced mixed gastric cancer.Finally,about 90%of the samples were identified in the verification set.The accuracy,recall rate and F1 value of the model were 92.34%,98.45%and 0.9551 respectively,which indicated the high reliability of the model.Especially in the positive classification,the ROC curve was further used to verify the model,and the measured AOC area was 0.91,indicating that the model performance was relatively reliable.By creat threshloder in qupath software,different thresholds are set and pixel is set to full.By comparing with raw data and surrounding tissues,undifferentiated components in mixed gastric cancer are separated,and smoothing sigma is adjusted according to threshold to distinguish tumor tissue from normal tissue.Finally,the model was evaluated by the comparison before and after segmentation.According to the results of the comparison chart,the model can segment the undifferentiated components in advanced mixed gastric cancer.Combined with the output values of the above two models and the outlined ROI region,the ratio of differentiated components to undifferentiated components in advanced mixed gastric cancer was quantified,and the information such as age,gender,tumor stage and tumor size of patients was calculated to analyze the relationship with LNM.ConclusionThe segmentation and recognition of high-grade components in MT can be achieved through the U-Net neural network architecture,while the recognition of undifferentiated components in MT can be achieved through QuPath.Ultimately,it was confirmed that there is a non-linear relationship between DUR and LNM,and DUR is an independent risk factor for LNM.The highest probability of LNM occurrence was found in the interval of 0.683 to 1.03 for DUR values.