Deep Learning-based Model For Predicting Muscle Invasive Of Digital Pathological Images Of Bladder Cancer

Objective:Accurate assessment of the invasion of the muscularis propria of the bladder is a necessary prerequisite for the pathological grading,staging diagnosis and prognostic evaluation of bladder cancer.Due to the subjectivity of pathological diagnosis and the complexity of image information,it is sometimes challenging for pathologists to distinguish non muscle-invasive bladder cancer(both Ta and T1 stages)from muscle-invasive bladder cancer(T2-T4 stages)based on microscopy.The aim of this study was to predict the region of cancer tissue and smooth muscle tissue in digital pathological images of bladder cancer patients using the Res-UNet deep learning segmentation model,and then to achieve accurate classification prediction of non-muscular invasive bladder cancer and muscular invasive bladder cancer using the Efficient Net V2 classification model.Methods:This study collected a total of 741 H&E-stained tissue sections,which included 419 muscular invasive and 322 non-muscular invasive from 350 patients(including 148 muscular invasive bladder cancer and 202 non-muscular invasive bladder cancer)who were pathologically diagnosed with bladder cancer after radical total cystectomy or transurethral electrosurgery of bladder tumor in the First Hospital of Shanxi Medical University from 2018 to 2021.These sections were scanned by scanner for whole slide image(WSI).All datasets were randomly split into training(70%)and testing sets(30%).First,49 WSI of muscleinvasive bladder cancer and 41 WSI of non muscle-invasive bladder cancer were randomly selected from the dataset as the segmentation model.Based on the Res-UNet network architecture,a digital pathological image segmentation model is established.The model uses the pathological image patches of manually labeled cancer foci and smooth muscle regions as input.The three types of background area,smooth muscle and cancer in the pathological image are segmented and trained to obtain the corresponding probability,and the effect of model segmentation is evaluated by calculating the DICE value.The model with the highest DICE mean was used as the optimal model,which was used to segment all the slides of the data set in the study.The predicted probability map generated by the segmentation model was concatenated with the original image as the input of the Efficient Net V2 classification model for classification prediction of muscle-invasive bladder cancer and non-muscleinvasive bladder cancer.The model training adopts the five-fold cross-validation method,and the area under curve(AUC)is calculated for all models in the cross-validation set.The model with the highest mean is used as the optimal model for the validation of the testing set.Finally,the average AUC of the five models is used as the predicted result.Results:Res-UNet segmentation model was used to verify that the average DICE of smooth muscle tissue and cancer tissue were 0.704 and 0.662,respectively.The average DICE of smooth muscle tissue and cancer tissue were 0.716 and 0.703,respectively.The segmentation performance was evaluated by analyzing the overlap rate between the predicted probability map and the ground truth,and the results showed that the model could accurately distinguish most of the tumor foci,smooth muscle regions,and background regions in the digital pathology images of bladder cancer.When Efficient Net V2 classification model was used to classify muscle-invasive and non-muscle-invasive bladder cancer,the classification model based on the predicted probability map combined with the original pathological image as input achieved the accuracy rates of 83.5%,82.8%,and 79.0%in the training set,validation set,and test set,respectively.And AUC reached 0.899,0.887 and 0.836,respectively.By contrast,the classification model using the original pathological image as input alone achieved the accuracy rates of 72.8%,79.8% and 77.0% and AUC values of 0.800,0.775 and 0.823,respectively.It was shown that the classification model can be used to differentiate between muscle-invasive and non-muscle-invasive bladder cancer.Furthermore,using segmentation prediction results in a classification model can improve the prediction performance of classification.Conclusion:The Res-UNet segmentation model constructed in this study can effectively segment cancer foci and smooth muscle tissue in WSI of bladder cancer.The segmentation results can help improve the performance of the subsequent classification model,and realize the accurate classification of muscle-invasive and non muscle-invasive bladder cancer.Thus,it can accurately distinguish Ta,T1 and ≥T2 bladder cancer and improve the accurate evaluation of p TNM staging by pathologists.Further development can provide a new auxiliary diagnostic tool for pathological practice,thereby accurately predicting prognosis and guiding clinical decision-making.