Study On The Assisted Classification Of Follow-up Knowledge Based On Low-dose CT Of Lung

With the follow-up work in the medical field gradually being valued,in the field of medical image processing,the task of obtaining follow-up information through relevant analysis methods has become more and more important.As far as lung cancer screening is concerned,in the early stage of diagnosis,doctors often need to evaluate the current CT screening results according to the classification report and give recommendations for follow-up screening.With the passage of follow-up time and number of follow-ups,it is easy for physicians to judge the results of follow-up screenings to ignore time information and related follow-up information,resulting in misdiagnosis and missed diagnosis.At the same time,in computer-aided diagnosis methods for lung cancer,most lung nodule classification models do not take into account the similar follow-up factors mentioned above,and there are often methodological defects in dealing with such problems.Around the background of low-dose CT long-term screening work,research has been conducted on the task of follow-up knowledge-assisted classification,and the improvement of the data imbalance caused by the missing problems in the process of collecting follow-up data during the long-term screening process has been carried out.The main search is as follows:First,research and analyze the distillation method used for follow-up knowledge to assist the classification of lung nodules.Inspired by the idea of knowledge transfer in the knowledge distillation network based on activation tensor attention,the task of long-term lung nodule classification is processed.The network structure of the knowledge distillation,the attention extraction method and the loss function are respectively improved.In detail,the specific activation tensor attention network uses a two-branch network to extract the strong knowledge into the weak knowledge network,and strong and weak knowledge are not distinguished by the multi-term distillation network,it is designed as a three-branch network structure according to different data periods.In the attention extraction method,this paper focuses on the attention analysis of the image edge features,size features,and nodules and solid features.In the loss function part,the network extracts both the output layer vector and the attention layer vector in distillation loss for training.The result of multi-period distillation network show that the classification accuracy was improved compared to the single-period network structure.Second,research and analyze the method of using the optimized meta-learning algorithm to improve the classification effect under the problem of data imbalance.In the actual process,there is an obvious imbalance in the low-dose CT screening data of multiple terms on an annual basis.This phenomenon causes a large difference in the data distribution of the input part of the multi-term distillation network,which affects the classification generalization ability of each branch network.In this regard,a method of using regularization function to solve the problem of data distribution imbalance in the optimization meta-learning algorithm used by the study to improve the algorithm structure and apply it to the multi-term distillation network,and at the same time,the comparative analysis in the experimental part shows that the use is different The L2 paradigm of the regular method has the best impact on the effect of improving generalization ability,and with the amount of data,the effect of the model on the improvement of generalization ability is 100,200,400,800,and the accuracy of 10 to 20 percentages is reached under the unbalanced data volume..