| Medical imaging requires doctors to distinguish the relationship between lesions and the severity of disease in the graph by human visual effects.The discrimination and segmentation of images directly affect the diagnosis results of doctors.However,with the emergence of a large number of new diseases and new medical images,the image complexity is increasingly high,often resulting in doctors cannot directly judge by the naked eye,but must rely on machine learning and other computer artificial intelligence technology for assistance.Therefore,more novel and practical machine learning algorithms are needed in the field of clinical imaging diagnosis.In this thesis,a new algorithm of deep learning is applied to recognize and segment medical images of heart,brain and lung.These include CT images of intracranial hemorrhage,identification of COVID-19 chest X-ray images,and segmentation of short axial MRI images of the left ventricle of the heart.The main purpose of this study is to provide reference for hospital clinical diagnosis of the above-mentioned diseases,and to partially expand the development of AI-assisted diagnosis and treatment in the field of medical imaging.This thesis mainly includes three parts:(1)Res Net18 and Dense Net121 models were selected to collect 2254 images and 514 normal CT images of five subtypes of intracranial hemorrhage(EDH,IVH,CPH,SAH and SDH)respectively.The data set was randomly selected according to 80%,20% and 20%,and divided into training set,validation set and test set.Transfer learning was applied to identify five subtypes of intracranial hemorrhage and normal CT images.(2)A total of282,336 and 371 chest radiographs of COVID-19,pneumonia and normal were collected.60%,10% and 30% were randomly extracted as training set,validation set and test set,respectively.AMRes Net model was used to identify COVID-19,pneumonia and normal chest radiographs.Accuracy,sensitivity,specificity,ROC curve and AUC value will be used to evaluate the performance of the model.The class activation diagram(CAM)was used to evaluate the interpretability and reliability of AMRes Net model.(3)200 short axial MRI images of the left ventricle of the heart were collected,and the segmentation network improved in this study was used to segment the cardiac lumen region of the atrium and ventricle.Five images were randomly extracted as the test set,and the intersection ratio(IOU)was used to evaluate the segmentation effect.Results were achieved in all three works: in the five subtypes of intracranial hemorrhage and normal CT image recognition,the overall accuracy of Res Net18 model and Dense Net121 model in the test set was 89.64% and 82.5%,respectively.The sensitivity of Res Net18 model to EDH,IVH,CPH,SAH,SDH and normal CT images was 98%,85%,80%,81%,93%,100%,respectively.The specificity was 88%,91%,91%,91%,89%,87%,respectively.The sensitivity of Dense Net121 model to EDH,IVH,CPH,SAH,SDH and normal CT images was 86%,73%,76%,81%,85%,98%,respectively.The specificity was 81%,85%,84%,83%,82%,79%,respectively.In the identification of COVID-19,pneumonia and normal chest radiographs,the overall accuracy of AMRes Net model in the test set was 94.9%.The sensitivity was 94%,90.5% and 92.2%,respectively.The specificity was 91.3%,93.1% and92.1%,respectively.In the heart cavity region of dividing atrium and ventricle,the cross ratio of 5 test images were 0.8148,0.8770,0.9077,0.8952,0.8570,respectively.Therefore,the deep learning method described in this thesis can accurately recognize and segment cardiac,brain and lung medical images.Res Net18,Dense Net121 and AMRes Net models have high sensitivity and specificity in CT images of intracranial hemorrhage and chest radiograph images of COVID-19,providing high reference value for clinical diagnosis of the above diseases.The improved segmentation network also achieves good segmentation effect in the short axial MRI image dataset of left ventricle. |
