Intelligent Determination Of Lymph Node Metastasis Of Papillary Thyroid Carcinoma Based On Multi-task Learning

In recent years,thyroid cancer has become the endocrine tumor with the highest incidence in the world,among which papillary carcinoma of thyroid is the most common pathological type,and is prone to lymph node metastasis in the early stage of the disease,which seriously threatens the health and safety of patients.However,as the most common imaging means for the clinical diagnosis of thyroid diseases,ultrasound has a low detection rate of abnormal cervical lymph nodes.Especially,due to the blockage of the sternum,lymph node metastasis in the central region of the neck is often missed,which brings difficulties for the accurate staging of preoperative diagnosis,resulting in some patients having to accept invasive treatment programs such as preventive lymph node dissection.Based on the above situation,this thesis combined the object detection task and classification task in the field of deep learning to establish an intelligent model based on multi-task learning to predict the probability of central lymph node metastasis in the occurrence of papillary thyroid carcinoma,so as to assist in guiding clinical diagnosis.The main achievements of this thesis are as follows:First,unified image labeling standards were established,which were annotated by professional doctors,and the experimental dataset of thyroid ultrasound images in this thesis was constructed.The dataset labeling included the task of nodular lesion detection and the task of central lymph node metastasis classification,with a total number of thousands of cases.Secondly,Libra R-CNN will be used as the nodule lesion detection network,based on which an automated way will be used to obtain the regional feature map of nodule lesions.Combined with the central lymph node metastasis classification task,an effective multi-task model will be built.Compared with single-task models,this model reduces the computational complexity,saves the storage space and improves the training efficiency.In the testing process,the model can get the results of two tasks by reasoning at one time,which improves the test efficiency.The experimental effect on the test set also proves the effectiveness of the model.Thirdly,this thesis introduced global image features including capsular membrane to supplement the features of local nodular lesions,and adjusted the weight of central lymph node metastasis classification task in the training process in stages to optimize the multi-task learning effect.Experimental results show that the MTL-GlobalW model proposed in this thesis can accomplish two tasks at the same time only by adding a small number of parameters on the basis of the lesion detection network,and the effect of the model is significantly better than that of a single task,which can efficiently detect nodular lesions and classify and predict central lymph node metastasis.