Research And Implementation Of Intelligent Cervical Lymph Node Detection Method Based On Multimodal Data Fusion

Cervical lymphadenopathy is a common cervical lymph node disease.If bilateral lymph nodes are invaded,the 5-year survival rate of patients is only 25%.There are usually subjective errors in clinical medical examination,and the diagnosis is time-consuming.Recently,more and more researches focus on the use of machine learning method to identify cervical lymph node disease.However,the ability of gray-scale ultrasound imaging technology to distinguish different tissues such as blood vessels and lymph nodes is poor.Only using gray-scale ultrasound images often results in the confusion of blood vessels and lymph nodes.Doppler ultrasound can provide blood flow signals inside and around lymph nodes,and can be used as a supplement to gray-scale ultrasound to distinguish blood vessels.At the same time,lymph nodes can be divided into normal,reactive and malignant lymph nodes.Reactive lymph nodes are a special kind of lymph nodes.There is little work on the identification of reactive lymph nodes.Therefore,this paper collected 1054 groups of gray-scale ultrasound and Doppler images for research,and proposed two model structures to solve the above two problems.In order to solve the problem of confusion among lymph nodes,blood vessels and muscles when using gray-scale ultrasound to locate lymph nodes,a multi-modal semantic segmentation model based on gray-scale ultrasound and Doppler is proposed.By introducing Doppler ultrasound image into the input to develop blood flow,the recognition result of this model is better than that of the single-mode model.The average dice value is 72.8%,and the single type dice value of lymph nodes is 78.6%.It can be seen that the model can distinguish lymph,muscle and blood vessels,and solve the confusion problem of the three to a certain extent.In traditional studies,only the normal/abnormal lymph nodes were distinguished.In this paper,the abnormal lymph nodes were further divided into reactive and malignant lymph nodes.Malignant lymph nodes need to be resected,while reactive lymph nodes need not be resected.In this paper,we first determine the distribution of normal,reactive and malignant in the data by collecting the results of biopsy,which is the gold standard of pathology.Then we propose a lymph node classification model based on the application of multimodal data and multi task learning network.The model takes ultrasound and Doppler images as the input,and lymph node status and related clinical indicators as the multi task output At the same time,the clinical indicators predicted by the model can be used as interpretable results to help doctors infer the types of lymph nodes.The accuracy of the model is 84.9%and 94.3%respectively.