Study On The Model Of Automatic Detection Of Early AIS Small Lesions On NCCT Based On DWI Using Yolov5 Network

Objective AIS(Acute Ischemic Stroke)has a narrow therapeutic time window,and early diagnosis and treatment of small brain ischemic lesions if they can be promptly diagnosed will greatly improve the outcome of AIS patients.Cranial DWI(diffusion-weighted imaging)is the most accurate test to diagnose small ischemic foci in the early stage of AIS.Emergent plain CT,as the first test of choice for AIS patients,is not sensitive for the diagnosis of early small ischemic foci,and is affected by factors such as physicians’ own experience and the consistency of AIS diagnosis is poor.AI(Artificial Intelligence)technology has the advantages of high efficiency,accuracy and repeatability,and has been widely used in the field of medical imaging.Previous AIS lesion detection models using the difference of HU(Hounsfield unit)between the left and right cerebral hemispheres with reference to CTP(CT perfusion)as the gold standard have verified the effectiveness of AI models for AIS diagnosis to a certain extent,but the ability to detect early small area ischemia is weak,the main reasons are as follows: In the early stage,the abnormal low density of ischemic lesions was not significant,the expression ability of small area of ischemic lesions was weak,and the image resolution of CTP was low.Therefore,in this study,based on the Yolov5 target detection deep learning network with improved loss function,using timely DWI(interval time between CT and DWI < 2h)images,an AI model that can accurately detect early AIS small-area ischemic lesions(stroke onset time < 6h,lesion diameter < 1.5cm on DWI)on CT images was constructed.Materials and Methods 1.Subjects: after collating the data according to the inclusion and exclusion criteria,275 AIS patients were finally included,including emergency plain CT scan images,MRI(magnetic resonance imaging)examination images,and related clinical history data.2.Data set construction: the cases were randomly divided into training group,validation group and test group according to the number of samples was 9:1:1.Plain CT images of the training and validation cohorts: the timely DWI images were fused with the plain CT images for image registration,and after the lesion area was identified,it was marked with a label on the original CT images.The CT images of the training group were processed and used for the training of the model,which was named the training set.The CT images of the validation group were processed to verify the detection accuracy of different loss function networks,and the optimal loss function was selected to construct a model,which was named the validation set.The unregistered and labeled original CT images of the test group were used to test the detection performance of the model on other samples,named the test set,and the test set data were used to compare the detection performance of the model with that of the manual group.3.Establishment of gold standard: in the validation set,the gold standard was the label of the validation set.In the test set,the gold standard is to mark the lesion area after the fusion of DWI and CT image registration and determination of the lesion area.DWIASPECTS(DWI image-based ASPECTS)was also calculated by trained physicians.4.Model construction: Yolov5 is used as the baseline model,different bounding box loss functions are used to train the verification network,and the optimal loss function is selected according to the detection result parameters of the validation set for storage.The M-CIo U(Modified CIo U loss)function in this paper has the best performance.Therefore,this study uses the Yolov5 deep learning network established by M-CIo U to obtain the final AI model after training.5.Model evaluation criteria: In the validation set,the optimal test results of deep learning network detection based on different loss functions are calculated,and the parameters include: m AP@50(Mean Average Precision),Recall and Precision.In the test set,calculate the specificity,sensitivity and accuracy of the final AI model;Conduct consistency analysis between CT-ASPECTS of model(CT-ASPECTS based on model test results)and DWI-ASPECTS.At the same time,in the test set,the doctor makes diagnosis on the CT scan image,obtains the specificity,sensitivity and accuracy of the artificial group diagnosis,and analyzes the consistency between CT-ASPECTS of manual group(CT-ASPECTS based on the artificial group detection results)and DWI-ASPECTS.6.Statistical methods: The consistency analysis between CT-ASPECTS of model and CTASPECTS of manual group and gold standard: ICC(intraclass correlation efficient)and weighted kappa test were used.The sensitivity,specificity and accuracy of the model and artificial group in each brain region were calculated.Results 1.Detection performance of different loss function networks in the validation set: M-CIo U(m AP@50 =0.7851;Precision=0.8237;Recall=0.8101)is optimal;CIo U(m AP@50=0.7060;Precision=0.6160;Recall=0.7377);DIo U(m AP@50=0.4629;Precision=0.8014;Recall=0.5942).2.The detection performance of the model in the test set: the specificity,sensitivity and accuracy of the model(98.87%,75.86%,96.20%)were higher than those of the artificial group(95.02%,63.79%,91.40%).3.Consistency test between CT-ASPECTS of model and CT-ASPECTS of manual group and DWI-ASPECTS: the consistency between CT-ASPECTS of model and DWIASPECTS ICC=0.669(95% CI 0.380-0.839,P<0.001);The consistency between CTASPECTS of manual group and DWI-ASPECTS ICC=0.452(95% CI 0.077 ~ 0.715,P=0.010);The weighted kappa coefficient of CT-ASPECTS of model and DWIASPECTS=0.447(95% CI: 0.255-0.699,P<0.001);The weighted kappa coefficient of CT-ASPECTS of manual group and DWI ASPECTS was 0.247(95% CI :-0.017-0.510,P=0.054).Conclusion Based on the improved Yolov5 deep learning network of loss function,this study constructs an AI model that can detect early small area lesions of AIS on CT plain scan images.The model showed excellent detection performance and general applicability,and the detection results were closer to the lesions indicated on DWI images than the physician group,and at the same time,it can provide assistance in the clinical assessment of AIS severity.