Detection And Localization Of Intracranial Hemorrhage Based On Deep Learning

Intracranial hemorrhage(ICH)refers to bleeding that occurs inside the cranium,which is a serious brain disease with high disability and mortality rate.It needs fast and accurate diagnosis to determine the location and subtype of intracranial hemorrhage,so as to improve the survival rate of patients.Brain computed tomography(CT)imaging has high accuracy and is widely used,which is the first choice for the preliminary diagnosis of intracranial hemorrhage.Experienced radiologists can diagnose intracranial hemorrhage and bleeding subtypes by patient’s brain CT images,but the diagnosis of intracranial hemorrhage by radiologists is very time-consuming,and the experience of physician has great influence on the accuracy of diagnosis.In this thesis,aiming at the problem of high missing rate and false detection rate of artificial diagnosis of intracranial hemorrhage,we carried out the detection and localization of intracranial hemorrhage based on deep learning.The main research work is as follows:1.Construction of CNN detection model based on window setting optimization.According to the characteristics of DICOM format brain CT images presenting different types of tissues and lesions in different HU ranges,combined with the window adjustment methods used by radiologists for diagnosing intracranial hemorrhage in clinical stage,the brain CT images under different windows were analyzed and compared,and the window setting optimization module was added as the input of CNN detection model.The experiment shows that the CNN detection model with window setting optimization module has better performance in the detection task of intracranial hemorrhage,and avoids the problem of image information loss caused by format conversion.In addition,the parameters of window setting optimization module can be used to find the best window display of intracranial hemorrhage images,and enhance the significant visibility of intracranial hemorrhage in CT images.2.Detection of intracranial hemorrhage based on improved RetinaNet.In view of the high missing rate and false detection rate of artificial diagnosis,and the low positioning accuracy of the lesion caused by the complex structure of the bleeding lesion area and the large difference in the morphology of different subtypes of bleeding lesions in the traditional object detection algorithm,the RetinaNet was adopted as the basic detection network framework of intracranial hemorrhage detection task,and the original RetinaNet model was improved from the aspects of basic feature extraction network structure,feature pyramid,loss function and training tricks.Experiments show that the improved RetinaNet model effectively reduces the missing detection rate of lesion detection and improves the positioning accuracy of lesion.3.Weakly supervised bleeding lesion localization based on multiscale feature fusion and Convolution Block Attention Module(CBAM).In view of the lack of highstandard labeling data of intracranial hemorrhage lesions in real scenes,the conventional fully-supervised object detection algorithm will be over-fitted due to the lack of data sets.Based on the idea of class activation mapping(CAM),combined with multiscale feature fusion and CBAM,a weakly supervised learning model based on classification network was constructed,which by bleeding subtype class label locating bleeding lesions.Experiments show that the weakly supervised learning model combined with multiscale feature fusion and CBAM increases the integrity of heatmap and ensures the classification performance.