Machine learning distribution model and auxiliary diagnosis system based on pathological sections from multiple parts of human beings

This study aimed to build up a machine learning model through resnet34 framework with strong generalization ability and an auxiliary diagnosis system to assist physicians to classify pathological sections from multiple parts of human beings. Usually, a pathological section can be distinguished between tumor, stroma, complex, lymph, debris, mucosa, adipose, and empty. U-net, resnet18, and resnet34 were employed to build the models. Accuracy was applied to access the performance of models since our data sets are homogeneous. Resnet34, selected to be our final framework, achieved the best performance with the highest accuracy between 95.0% and 97.2% with training on the data sets, adjusting options, and adding a fully connected layer. Since it is not possible to deploy the required machine learning environment on every physician's computer, we deployed an online recognition system based on the models, presented as a website, for physicians to use.