| Stroke is currently an important cause of disability and death in adults domestically.In recent years,the prevalence of ischemic stroke,an important type of stroke,has increased year by year.And large hemispheric infarction（LHI）in the hyperacute stage,as a more serious type of ischemic stroke,is characterized by sudden onset and rapid deterioration of the disease,for which rapid diagnosis and thrombolysis / thrombectomy are required for the benefit of the patients.If the patients can be diagnosed quickly and accurately during the ultra-acute period with the degree of ischemia graded,we can carry out accurate intervention treatment according to the situation of the patients.Computed tomography（CT）has become the first choice for rapid diagnosis and evaluation of stroke patients due to its advantages of fast imaging and low cost.Therefore,we hope to provide personalized thrombolysis / thrombectomy for patients with ischemic stroke through computer-aided technology combined with CT images,so as to solve the problems of under treatment and over treatment in clinic.First of all,we must realize the automatic classification and identification of LHI.According to the characteristics of ischemic stroke CT images in the ultra-acute stage,we propose an intelligent diagnostic scheme of “segmentation before classification”,in which two major tasks are included: intracranial ischemic regional segmentation and the LHI classification.In terms of intracranial ischemic region segmentation,considering the complexity of the task,traditional image processing methods are difficult to meet the needs,while the emergence of deep learning theory and technology provides a new path for the precise segmentation of ischemic areas in hyperacute phase.Therefore,this paper proposed a new segmentation network model CF-Net,which can realize the end-to-end segmentation of CT two-dimensional images.The overall structure of the network adopted the form of “encoding and decoding”,with the encoder used as the feature extractor and the decoder as the pixel classifier.In the network,we used the Non-localmodule to extract bilateral differences of the brain and improve the accuracy of segmentation.Meanwhile,we also used the channel information fusion module and the spatial information fusion module to strengthen its sensitivity to the main feature information.And because the edge of ischemia is not obvious in CT,we used Tversky loss function to increase the recall rate of the network.We experimented in 50 patients with hyperacute ischemia,with the Dice Similarity Coefficient（DSC）and the intersection over union（IOU）finally reaching 90.36% and 86.75% respectively.For the classification algorithm,we compared the classification method based on ischemic tissue volume and the recognition results of different networks based on 3D classification.According to the clinical diagnosis of hyperacute LHI,this paper adopted the volume of ischemic tissue greater that 71 ml as the classification criterion for LHI.At the same time,we also verified the classification effect of segmented ischemic tissue using three kinds of 3D convolution networks: 3D-VGG,3D-Res Net and 3D-SENet.The method was verified by data collected from 135 plain CT images,including 10 healthy volunteers,40 patients with LHI and 85 patients with non LHI cerebral infarction.The experimental results show that the classification method based on volume has the best recognition effect,with the accuracy rate reaching 91.67%. |
PDF Full Text Request