Study On The Detection Of CT Reports Of Tuberculosis Based On Transfer Learning

Today,when the new coronavirus has not completely subsided,the diagnosis and treatment of other kinds of diseases（such as tuberculosis）is still an issue that cannot be ignored by experts in this field.Computer tomography is an important method to diagnose tuberculosis.Due to the complexity and particularity of CT images,the CT report is mainly done by the radiologist.The development of artificial intelligence has led to the acceleration of the widespread application of deep learning in tuberculosis treatment.How to efficiently identify the pulmonary tuberculosis lesions on the patient’s CT is extremely important to help doctors quickly make a diagnosis.The attention mechanism is widely used in the field of image processing,Therefore,the attention mechanism is also a main method in the field of medical images,such as CT image classification.Challenges due to the nature of 3D data: variable volume size,GPU exhaustion during optimization.The independent processing of a single slice in the two-dimensional CNN deliberately discards the depth information,resulting in poor performance of the expected task.We solve the above problems through data preprocessing.One method is to preserve the depth information by converting the three-dimensional CT image into a multi-axial（axial,sagittal,and coronal）two-dimensional projection.For the study of CT reports of tuberculosis,this article conducts research in two parts.The first part of the study is to score the severity of tuberculosis.This paper proposes an improved version of the VGG16 model for this task,and uses a machine learning classifier to make predictions in combination with patient metadata information.The second part of the research is the automatic generation of lesion labels for tuberculosis CT reports.This paper proposes a convolutional module attention CNN integrated model under the transfer learning method.The model mainly adjusts the model performance based on the transfer learning method combined with the convolution module attention mechanism and the Ghost module on a series of processed data sets.To the best,then to the test data set to predict the result.On the ImageCLEF019 tuberculosis severity score dataset,our model achieved an AUC of 0.878,compared with directly using CNN to process two-dimensional slices,the effect is improved by about 7%.On the Image CLEF2020 pulmonary tuberculosis CT report lesion label automatic generation task data set,the model used in this thesis reached mean＿auc 0.723 and min＿auc 0.647 on the test set,respectively.After using the attention mechanism of the convolution module,the model reached mean＿auc 0.792 and min＿auc 0.693 on the test set,respectively,which proved the effectiveness of the improved model proposed in this thesis in the task of generating CT report lesion labels.