Deep Learning Method Based on CT Images to Predict the Pathological Differentiation of Intrahepatic Cholangiocarcinoma

Intrahepatic cholangiocarcinoma (ICC) is an adenocarcinoma that originates from the epithelium of the secondary bile ducts and their branches. In recent years, the incidence has shown an increasing trend. As an important clinical indicator of ICC, the degree of pathological differentiation is of great significance for clinical treatment. In this work, we retrospectively analyzed the clinical data of 443 patients with medical follow-up tests from West China Hospital in the past, and the cases were randomly divided into trainset and testset, which both included cases labeled as moderate to high differentiation or low differentiation. The high-dimensional features of the images were extracted by a deep convolutional network (DCNN), where a re-weight method was used to mitigate the impact of the class imbalance prevalent in the medical dataset in the learning of the deep convolutional network, and finally the predicted probability of the degree of differentiation was output by a fully connected network. In this work, we used the Re-weight method and used the mean statistical patient differentiation degree probability and finally predicted the differentiation degree of ICC patients using model integration. We divided the differentiation degree into two levels, low and medium-high differentiation degree, focusing on the accuracy of the lower differentiation degree. We divided the degree of pathological differentiation into two levels, namely, low and medium-high differentiation. The test set is randomly divided, where the category ratio is 1:1, for more objectively verify the performance of the model. By the method proposed in this paper, a final accuracy of 72% was obtained, which is an 8% improvement compared to 64% in baseline, aiming to better guide the clinical treatment of patients.