Research On Key Algorithom Of Detection Of Tuberculosis In Radiographs

Tuberculosis is the second leading cause of death in the world,ranking first in infectious diseases.According to the global report issued by the World Health Organization(WHO),the death of most TB patients can be prevented by initial and appropriate treatment.But unfortunately,the cost of most current diagnostic methods is not allowed to be used on a large scale in developing countries most affected by tuberculosis.The diagnosis of tuberculosis is still a major challenge.In current tuberculosis analysis methods,radiographs are fast,convenient,and inexpensive,and are often used as a means of extensive census.With the increase in the amount of medical data,it is increasingly difficult for radiologists to maintain the same diagnostic efficiency of radiographs for all patients,and there is an urgent need for high-precision computer-aided diagnostic systems to help radiologists maintain diagnostic quality.This paper first studies the basic structure of convolutional neural networks and its application in target detection tasks.At the same time,the experimental data of tuberculosis is introduced.Considering the particularity of medical images,tuberculosis radiographs are pretreated to enhance the lung characteristics.Next,the paper designs the automatic detection algorithm of tuberculosis based on deep learning.The basic detection network framework adopts RetinaNet model to realize multi-scale detection.Then,for the problem of high false negative rate of the original model,the Anchor structure parameter initialization method,AnchorOriented algorithm and the new difficult learning sample mining loss function are proposed to optimize the original model from the perspective of network structure and parameters,effectively reducing the false negative rate and improving detection accuracy.Then,a high-precision classification network of pulmonary tuberculosis is designed by combining the chest segmentation information,and the design is added to the detection network as a false-positive constraint.The tuberculosis classification network consists of a U-Net-based chest segmentation path and feature extraction path,global average pooling operations and the final full convolutional layer to implement class prediction.The global information and local information of the chest radiograph are effectively utilized by using the complete chest radiograph and the lesion image block as the classification network input.Finally,in order to reduce the false positive rate of the model without changing the false negative rate,the designed high-precision tuberculosis classification network is added to the detection network as a false positive constraint.After detecting the network to obtain the prediction result,this paper designs an effective false positive constraint algorithm to obtain the final tuberculosis detection by integrating the output of the detection model and the classification model.The tuberculosis detection model designed in this paper achieved an accuracy of approximately 92% on both test data sets.