Prediction Method Of Mri Pancreatic Cancer Lymph Node Metastasis Stage Based On Radiomics And Deep Learning

Pancreatic cancer has an insidious onset and atypical early symptoms,and its five-year survival rate is only 8.5%.Currently,the only way to treat pancreatic cancer is to perform surgical resection.The staging of lymph node metastasis can only be determined by postoperative pathological sections.An important factor affecting the prognosis of surgery is lymph node metastasis.Magnetic resonance imaging(MRI),as a non-invasive examination,is widely used in the detection of pancreatic cancer.If the lymph node metastasis of pancreatic cancer can be predicted safely and accurately before operation,it will have great guiding significance for the diagnosis and subsequent adjuvant therapy of pancreatic cancer.The first part of this thesis is to predict lymph node metastasis staging in pancreatic cancer based on radiomics.Firstly,the radiomics features of two unenhanced and three enhanced MRI sequences are extracted respectively,and then the extracted single-sequence and multi-sequence features are screened and modeled using three feature selection methods and three different classifiers.Among them,the methods of feature selection are least absolute shrinkage and selection operator(LASSO),recursive feature elimination(RFE)and elastic network,and the classifiers are random forest,logistic regression and support vector machine(SVM).Finally,the clinical indicators provided by radiologists are added to the multi-sequence model to improve the prediction accuracy of pancreatic cancer lymph node metastasis staging.The second part of this thesis is to predict lymph node metastasis staging in pancreatic cancer based on deep learning.For each individual MRI sequence,a single branch feature extraction network(SBFENet)is proposed by the introduction of a segmented image of pancreatic tumors,and the design of a multi-level interactive spatial attention(MISA)module that utilizes the segmented image to guide classification and better learn tumor information by fusing feature maps of two different scales.In order to further improve the accuracy,a multi-branch feature fusion network(MBFNet)is proposed based on the single branch network,which embeds SBFENet into the multi-branch network.An interactive channel self-attention(ICA)module and a multi-sequence fusion(MSF)module are designed to allow the network to measure the importance of each channel in the same sequence and the correlation of tumor features between different sequences during the learning process.In addition,the radiomic features of the five MRI sequences and the clinical features provided by the radiologist are also added to the multi-branch network to improve the accuracy of the prediction model for lymph node metastasis staging.The dataset is collected from Department of Radiology,Changhai Hospital,Naval Medical University.A total of 356 patients with pancreatic cancer were included in the study with abdominal MRI.The combination of LASSO and logistic regression is used to model the radiomics and clinical features of the five MRI sequences,achieves 0.6798 and 0.7136 on ACC and AUC,respectively.By combining deep learning,radiomics features and clinical features to predict pancreatic cancer lymph node metastasis,MBFNet achieves 0.7303 and0.7286 on ACC and AUC,respectively.The results show that the proposed method can better achieve the task of preoperative prediction of lymph node metastasis of pancreatic cancer.