Developing And Verifying Automatic Detection Of Active Pulmonary Tuberculosis From Multi-slice Spiral CT Images Based On Deep Learning

Objective In many remote areas where tuberculosis is endemic,due to lack of radiologists and inexperienced doctors,it is a difficult task for radiologists to diagnose tuberculosis(TB)by using multi-slice spiral computed tomography(CT).In order to solve this problem,this research has developed an artificial intelligence(AI)-based automatic detection of active tuberculosis(ATB)system to help radiologists simplify the diagnosis process of ATB,reduce the radiologist’s workload and improve diagnostic accuracy.Date This study retrospectively collected CT images(slice thickness:5mm)of 846 patients(476 cases of ATB,150 cases of non-tuberculous pneumonia,and 220 cases of normal)from the Affiliated Hospital of Hebei University from April 2016 to May 2019.In this article,positive sputum smears are the gold standard for ATB patients,and CT diagnosis report results are the gold standard for normal and pneumonia patients.This study divided the data into a training group and a test group.The training group randomly selected 337 cases of ATB,110 cases of pneumonia and 120 normal cases,and the test group selected the remaining 139 cases of ATB,40 cases of pneumonia and 100 normal cases.The data of the training group and the test group are independent samples.Methods 1、Spss25.0 software was used for statistical analysis.The independent sample t test was used to compare the ages of the training group and the test group.Theχ~2 test was used to compare the genders of the training group and the test group.P<0.05 indicated a statistically significant difference.2、First,Label Img software was used to label 337 ATB lesions in the training group.Second,the U-Net segmentation network(U-Net)is used to train the data in the training group,and at the same time,the image segmentation training is performed on the labeled ATB area,and the U-Net model that automatically recognizes the ATB lesion is obtained.Third,in order to improve the accuracy of diagnosis,this article uses image processing methods for the CT layers diagnosed by U-Net as ATB lesions to screen out some layers misdiagnosed by U-Net,and make 2D ATB lesions in continuous CT images Converted to 3D lesions,when the patient meets all the above conditions,it is diagnosed as an ATB patient,so this AI diagnosis software is composed of U-Net network plus image processing methods.Finally,this study uses a test group to test the diagnostic performance of AI software.Results The age(t=1.167,P=0.243)and gender(χ2=3.708,P=0.054)of patients in the training group and test group.In this article,the AI diagnostic software’s evaluation indicators for the test group’s detection results are:ROC(Receiver Operating Characteristic)curve AUC(Area Under Curve)value is 0.980,accuracy(Acuracy,Acc)is 0.968,sensitivity(Sensitivity,SS)is 0.964,Specificity(SP)is 0.971,Positive Predictive Value(PPV)is 0.971,and Negative Predictive Value(NPV)is 0.964.The results show that this AI diagnostic software is useful for diagnosing ATB and differential diagnosis of non-ATB(Pneumonia,normal people)has high accuracy and performance.Conclusion This study successfully developed an AI software tool for automatic detection of ATB by chest CT.The results of this study show that the use of this AI diagnostic software can not only accurately diagnose ATB,but also lay a solid foundation for the next step of AI diagnosis ATB in clinical medicine.