Detection And Metastasis Prediction Of Pulmonary Mediastinal Lymph Nodes Based On Deep Neural Network

Lung cancer is one of the most common diseases and evaluation of the lymph nodes plays an important role in diagnosis and treatment of it.Enlarged lymph node is considered having a significant relationship with lymph node metastasis.Doctors evaluate the lymph node metastasis of patients by CT images and/or invasive methods of lymph node staging,and choose the corresponding treatment according to the evaluated results.Meanwhile,the detection of lymph nodes affects the patient’s disease development,disease diagnosis,selections of treatment and follow-up.However,due to the lack of medical resources and the large population of patients,reading CT images manually makes doctors have intensity of working and patients need to wait a long time for their diagnosis results.Meanwhile,due to these limitations,it is difficult to perform invasive mediastinal lymph node staging methods for patients to ensure the accurate preoperative staging of their lymph nodes.Moreover,due to the different format of CT images generated by different imaging equipment or different settings,it has an impact on the performance of lymph nodes detection in some degree.Therefore,it is necessary to develop solutions for the above problems.With the emergence of deep learning in natural images and medical images in recent years,CAD system based on deep learning is becoming more and more popular,and performance of it is becoming more and more considerable.The detection and evaluation of lymph nodes are important for the therapy of patient.In order to assist solving these problems,we developed a lymph node detection algorithm for CT based on deep CNN and a predictive model for NSCLC patients to assist non-invasive decision-making of N2 lymph node metastasis.For the detection of mediastinal lymph nodes,we constructed a lymph node detection algorithm for CT images based on deep CNN to assist evaluation and management of lymph nodes of patients for doctors.It contains three parts: a down-sampling encoder network,an upsampling decoder network and an RPN network.And it does not need the spatial priors of mediastinal tissue.Meanwhile,it adopts the method of stitching multi-scale features to alleviate the problem of small data.The evaluated result of it on NIH opensource dataset shows good performance in lymph node detection.To solve the problem of different formats of mediastinal lymph node CT images,we proposed a mediastinum and lung segmentation method to quickly segment the lung and mediastinum regions on CT images.For preoperative lymph node assessment,we analyzed the high-risk factors of N2 lymph node metastasis,and developed a preoperative non-invasive N2 lymph node metastasis prediction system based on neural network to provide the prior of the preoperative lymph node staging.