Medical Imaging And Analysis Of Acute Ischemic Stroke By Using Deep Learning

Stroke ranks as one of the leading death causes in China.It has the characteristics of high morbidity rate,high mortality rate,high disability rate,high recurrence rate and high economic burden.The proportion of acute ischemic stroke(AIS),also known as cerebral infarction,is more than 70%.According to the 2018 Chinese guidelines for diagnosis and treatment of acute ischemic stroke,it is strongly recommended that the suspected stroke patients should complete the basic evaluation of brain CT within 60 min after being admitted to hospital and receive the treatment within the 6-hour golden window.Therefore,AIS puts forward higher requirements for rapid and accurate medical imaging examination.Quickly determining the irrecoverable area of brain tissue function loss and the time since stroke(TSS)are two crucial aspects in the diagnosis of AIS.Interventional doctors should determine the treatment method of intravenous thrombolysis or mechanical thrombectomy according to the size of the infarct core and the TSS.CT perfusion(CTP)imaging is recommended by the Chinese AIS treatment guidelines for the evaluation of infarct areas.The cerebral blood flow(CBF),cerebral blood volume(CBV),mean transit time(MTT),time to peak(TTP)and other perfusion maps calculated from CTP are the basis for radiologists to estimate infarct volume.CTP imaging needs to scan the patient’s brain repeatedly for 30 times(30 passes)in 50 seconds.The repeated scans will greatly increase the radiation dose.It has great application value to ensure the reliability of the image while reducing the radiation dose of CTP imaging.Delineating the infarct core on CTP parameter maps manually is time-consuming and laborious for radiologists.Because there is no uniform threshold,the delineation may be inaccurate,which will affect the choice of AIS treatment.Patients with unknown TSS are excluded from thrombolysis since the thrombolytic therapy is strictly limited to the 4.5h time window of onset.For those patients,the clinical screening is usually performed by means of high-signal mismatch between diffusion weighted imaging(DWI)and fluid attenuated inversion recovery(FLAIR)images.However,the human-derived DWI-FLAIR mismatch model has low accuracy in predicting TSS.To resolve the above-mentioned problems of the high radiation dose of AIS CTP imaging,the long time consuming of infarct core delineation,and the low TSS prediction accuracy of human-derived DWI-FLAIR mismatch,we study the stroke imaging and analysis methods based on deep learning.This paper mainly focuses on the high-quality reconstruction of low-dose CTP imaging,the accurate prediction of infarct core based on CTP maps,the automatic identification of stroke patients within or without 4.5h using DWI and FLAIR features.Our main contributions are as follows:(1)A cascaded network is proposed to achieve the high-quality reconstruction of low-dose CTP imaging.Considering that CTP imaging needs to scan stroke patients 30 times in a short time,the method of reducing imaging tube current and reducing the number of repeated scans can be used to reduce the radiation to patients.However,reducing the imaging tube current will bring in a lot of speckle noise in the image domain,and reducing the number of scans will affect the calculation of perfusion parameters.These shortcomings will affect the reliability of perfusion parameter data.Here,we construct a cascaded convolutional network.The first sub-network eliminates the speckle noise caused by low tube current in the image domain,and the second sub-network restores the low-sampled images to 30 passes in the time domain.The experimental results based on the simulation data verify the effectiveness of the proposed cascaded network.Compared with the classical algorithm,the proposed algorithm can better preserve the structural details of brain image,and the processing speed of the proposed algorithm is much faster.(2)A multi-sequence semantic analysis network ISP-Net(Ischemic Stroke Prediction Network)is proposed to predict the AIS infarct core.Through the analysis of CTP parameter maps,the infarct core of patients can be predicted,so as to provide information for the choice of thromobolysis or thrombectomy.The inputs of ISP-Net are the native CTP,CBV,CBF,MTT,Tmax,and the outputs are the label maps of infarct prediction.Considering the diversity of infarct core location and size,a multi-scale convolution module(MSAC)fusing the image attention information is introduced to ISP-Net,which can effectively improve the accuracy of infarct core prediction.We carry out retrospective evaluation on a stroke dataset collected from multi-centers.The experimental results show that ISP-Net outperforms several advanced semantic segmentation models.ISP-Net has higher reliability than the threshold algorithm which is commonly used in medical workstations,demonstrating the excellent performance of infarct volume estimation.(3)An automatical method is proposed for identifying whether the TSS of stroke patients is within 4.5h.First,we develop a cross-sequence convolutional neural network to accurately segment the stroke lesions from DWI and FLAIR images.Second,the radiomics features are extracted from DWI and FLAIR according to the segmentation regions of interest(ROI).Finally,the features are fed to 5 machine learning models and identifying TSS ≤4.5h or >4.5h according to the voting of the 5 classifiers.The cross-sequence network proposed in this paper can integrate DWI feature information into the segmentation of FLAIR image,so as to effectively solve the problems of low contrast of infarct core signal and the interference by other tissue signals in FLAIR image.The voting of the 5 models can improve the robustness of classification results and avoid the problem of low sensitivity or specificity of a single classifier.We carry out retrospective evaluation on a stroke dataset collected from multi-centers.The experimental results show that the accuracy of this automatic method is obviously better than the human-derived DWI-FLAIR mismatch,illustrating the potential for automatic and fast TSS identification.