Research On Detection Of Intracranial Hemorrhage Based On Deep Learning

Intracranial hemorrhage is an urgent cerebrovascular disease that can lead to severe disability or death in humans,and accurate diagnosis and timely intervention programs by clinicians will greatly improve the survival rate of patients with hemorrhage.Currently,the use of computed tomography(CT)imaging is the preferred option for physicians to initially diagnose intracranial hemorrhage,and deep learning-based CT image assisted diagnosis study of intracranial hemorrhage can effectively reduce the rate of missed diagnosis and misdiagnosis by physicians.The computer-aided diagnosis of intracranial hemorrhage study mainly includes two parts: intracranial hemorrhage subtype classification and intracranial hemorrhage lesion area segmentation.In intracranial hemorrhage subtype classification,most studies are based on a single convolutional neural network,and these methods often ignore the correlation between hemorrhage CT slices,and the prediction probability with one output alone is not enough to guarantee its reliability.It is a challenging and difficult task to obtain deep learning models with higher reliability and robustness.In intracranial hemorrhage lesion region segmentation,V-Net-based lesion segmentation has shown its basic performance advantages,however,its own structural design still has the following problems: the large number of parameters in the original ordinary convolutional approach in V-Net leads to longer model training time;the simple structure of encoder and decoder,thus unable to extract deeper features and restore the edges and details of image lesions This limits the segmentation performance of the network.In order to solve the above problems,the research of this paper is as follows:1.For intracranial hemorrhage detection subtype classification,a joint CNN and RNN method is proposed for intracranial hemorrhage detection classification of CT scans.In the CNN module,the residual network Res Net101 is selected as the base network and the channel attention mechanism,the SE module,is added.The addition of the SE module gives the CNN network a more powerful feature extraction capability and allows the adjustment of the weights for each channel,thus making the model more generalizable.Considering the correlation of sequences between CT slices,a bidirectional gated recurrent unit GRU network is introduced,and the sensitivity of RNN to sequence data can be used to better process CT sequence data and improve classification accuracy.The feature vector is also reduced in dimensionality during the CNN transfer to RNN,thus reducing the model computation.The comparison experimental results show that the proposed method in this paper has the best performance in four evalution indexes:accuracy,AUC,sensitivity and specificity.2.For intracranial hemorrhage lesion segmentation,this paper improves the V-Net network model and applies it to the task of intracranial hemorrhage segmentation.Based on the V-Net network firstly,the normal convolutional approach is replaced by the deep separable convolution,which speeds up the training speed of the model.Then channel attention mechanism and hybrid attention mechanism are added to the encoder and decoder,respectively.The SE module is introduced in the encoder,which is designed to strengthen the feature extraction ability of the original network and increase the perceptual field of the feature map.The CBAM module is introduced in the decoder,so that the original network can adaptively adjust the weights between different channels in the feature map,thereby improving the performance of the model.The results of comparative experiments show that the improved V-Net performs best in all indicators,among which the DSC reaches0.732,which is 4.4% higher than the original V-Net.