Deep Learning-based Prediction Of Postoperative Recurrence Of Intrahepatic Cholangiocarcinoma

Intrahepatic cholangiocarcinoma is a highly malignant tumor that has seriously affected people’s lives for a long time.In the early stage,its symptoms are not obvious,so it is not easy to be detected,and when it is detected,it is mostly in the middle and late stage,and because of its special location of the lesion,it makes the treatment very difficult,and a serious problem is that its prognosis is very poor,and according to different studies,the 5-year survival rate is less than 30% among the patients who underwent surgery.Therefore,it is of great importance to go for the construction of a model that can accurately predict the postoperative recurrence.In order to efficiently make predictions for postoperative recurrence in patients with hepatobiliary cell carcinoma,thesis proposes a parallel mixed-model prediction framework based on convolutional networks and Transformer networks and a serial mixed model prediction framework based on reinforcement learning.Meanwhile,a high-quality version of the dataset consisting of 455 patients with 13728 CT images was constructed.A high-quality version of the dataset consisting of 455 patients with a total of 13728 CT images is constructed.Most of the traditional studies on recurrence after hepatobiliary duct surgery have used methods such as imagingomics and machine learning.But deep learning has long shown great advantages for image feature extraction.Therefore,a hybrid model of convolutional network and Transformer network is proposed in thesis.Through a parallel dual network structure,we make full use of the respective advantages of convolutional network and Transformer network to extract features to complete the prediction.At the same time,in addition to making full use of the CT images,the index data corresponding to the cases are fused with their image data to supplement the information that cannot be expressed in the CT images.Through continuous optimization training,a postoperative recurrence prediction model with good performance is finally obtained.Since a simple fusion of indicator data with image data cannot fully exploit its advantages,thesis also proposes a prediction model based on reinforcement learning,which gives different weights to indicator data.By maximizing the reward value,the weights of each indicator data are dynamically adjusted to make the indicator data work better.In thesis,the accuracy of the model on the test set data is used as the reward value,and the optimal weight model is obtained through continuous optimization,which also leads to a prediction model with good generalization for postoperative recurrence of intrahepatic cholangiocarcinoma.Extensive experiments have shown that the proposed Conformer model-based prediction framework and reinforcement learning-based prediction framework have excellent performance and generalizability.