Classification And Localization Of Intracranial Hemorrhage Subtypes Based On Deep Learning And CT Images

Intracranial hemorrhage is defined as bleeding within the skull.It is a serious and severe craniosynostosis,known for its high mortality and lethality rate.Patients usually need immediate follow-up diagnosis and identification of the location and subtype of the intracranial hemorrhage lesion in order to improve the chances of successfully saving the patient.Computed tomography(CT)is a highly accurate and widely used clinical tool for the initial diagnosis of intracranial hemorrhage symptoms.Generally,after a patient has undergone a CT scan,the physician will use the resulting CT scan images of the brain to make an effective diagnosis,but this process is very time consuming and tests the physician’s skill level,so the physician’s qualifications and professionalism often affect the outcome of the patient’s visit.Because of human error,there are often missed or incorrect diagnoses,so computer-aided diagnosis and treatment systems are important for physicians to diagnose and plan subsequent treatment.Most studies of intelligent-assisted diagnosis and treatment of intracranial hemorrhage have used CNN-RNN network architectures.These methods suffer from the following problems:(1)the currently used CNN networks,such as Res Net and Dense Net,have the disadvantage that they cannot adequately represent the spatiotemporal characteristics of CT cranial images;(2)the ratio of positive to negative samples in the data set about CT scan results is severely imbalanced;(3)problematic CT slices often have one or even multiple lesions but the correlation between lesions is usually ignored.To address the above problems,this paper classifies and localizes intracranial hemorrhage based on deep learning methods.The main research works are as follows.1.Vision Transformer-based subtype identification of intracranial hemorrhage.Aiming at the disadvantages of poor parallelism and long training time of existing RNN and LSTMbased intracranial hemorrhage detection algorithms,an end-to-end Vision Transformerbased algorithm model is proposed,which greatly reduces the training time of the model while considering the time-series features of cranial CT images.Meanwhile,under the DICOM format,the range values of Hu in the images obtained from CT scans will be different for different kinds of tissues and different lesions.On this basis,considering the window function that radiologists often resort to in clinical diagnosis,this paper studies and compares the brain CT images under different windows,and selects the pre-processed data after the best window setting as the input of CNN detection model to improve the accuracy of intracranial hemorrhage identification.2.Retina Net-based intracranial hemorrhage localization study.Because the lesion location structure of intracranial hemorrhage is complex and the morphology of the lesion varies greatly among different subtypes,the problem of lesion localization accuracy is not effectively solved.To address the above problems,this paper then starts from the feature extraction network structure,training techniques and Anchor settings in the original Retina Net model,and gradually optimizes them.Through comparison experiments,it can be found that the improved model has better results than the three target detection models,Faster R-CNN,Retina Net and YOLOv4,mentioned in Chapter 2.3.A study of intracranial hemorrhage localization based on weak supervision and CAM.A weakly supervised method for accurate hemorrhage localization using only location-free labeling of axial slices based on multi-instance learning is proposed.The method is able to find the specific coordinates of intracranial hemorrhage by heat map while ensuring interpretability.