The Study Of Identification Of Accelerated Craniocerebral Injury And Decelerated Craniocerebral Injury Based On Computed Tomography (CT) Imaging And Deep Learning Algorithm

Background and purposeIn the practice of forensic medicine,blunt force craniocerebral injury is the most complicated mechanical injury,among which accelerated craniocerebral injury and decelerated craniocerebral injury caused by linear motion are the most common.The scientific identification of accelerated craniocerebral injury and decelerated craniocerebral injury is of great significance for the speculation of the manner of injury and the instruments causing injury,the responsibility division of traffic accidents and the cause analysis of the falling injury.At present,CT technology,as an imaging tool,has been more and more applied in the field of forensic medicine.A large number of imaging data of craniocerebral injury can be obtained by CT scan,which can help to systematically analyze and understand the morphological characteristics and injury rules of accelerated and decelerated craniocerebral injury,and thus assisting in identify accelerated craiocerebral injury and decelerated craniocerebral injury and improving the scientificity and objectivity of the conclusion of appraisal.However,the identification of a large number of imaging data not only increases the workload of forensic experts,but also leads to the omissions of the minor injuries or inaccurate conclusions due to the uneven level of film reading of forensic experts.At the same time,relying solely on forensic experts for subjective empirical discrimination of imaging data will also lead to inconsistent cognition and lack of quantitative data support.In recent years,the development of deep learning algorithm is in full swing,which has been widely used in many fields,like image recognition,speech recognition,computer vision,and has achieved obvious achievements.In the field of forensic medicine,the combination of deep learning algorithm and imaging technology to achieve computer-aided imaging diagnosis has gradually become a research focus and an important direction.Therefore,this study intends to make use of the advantages and application of deep learning in image recognition to explore its feasibility in identifying accelerated and decelerated craniocerebral injury based on craniocerebral CT images,in order to establish an objective,efficient and accurate auxiliary imaging identification method for accelerated and decelerated craniocerebral injury.Methods1.The imaging data of 214 cases with accelerated craniocerebral injury and decelerated craniocerebral injury were collected retrospectively and the craniocerebral injury imaging database was established.Chi-square test of R×C contingency table was used to analyze the relationship between stress site and injury site in accelerated and decelerated craniocerebral injury,and P<0.001 was considered as significant.χ~2test was used to analyze the differences in stress site,injury distribution and common injury types of craniocerebral injury under different injury modes,and P<0.05 was considered as significant,thereby summarizing the morphological characteristics and general distribution rules of craniocerebral injury under different injury manners.2.The above craniocerebral injury imaging database was used as the experimental object,and the craniocerebral injury imaging data were divided into training dataset and testing dataset according to random sampling method.The training dataset was used to train the segmentation performance of the model,and the testing dataset was used to test the segmentation performance of the model.U-Net semantic segmentation model was constructed to study the segmentation of bleeding area(including scalp hematoma area and intracranial hemorrhage area)based on brain CT images,and Dice coefficient was used to evaluate the segmentation performance of the model.3.The craniocerebral imaging data of 130 normal cases were added to the craniocerebral injury imaging database and used as a control group.The craniocerebral imaging data were divided into training and validation dataset,testing dataset according to random sampling method.The training and validation dataset were used for model training and parameter optimization,and the testing dataset was used to test the classification performance of the model.Vgg16 model,Res Net18model and Inception＿v3 model,were constructed to classify the CT images of accelerated craniocerebral injury,decelerated craniocerebral injury and normal brain,respectively.The accuracy,precision,recall,F1-value and the area under receiver operating characteristic(ROC)curve(namely AUC)are adopted to evaluate the performance of the three models in the testing dataset.Result1.There was a strong correlation between the stress site and the injury site in accelerated craniocerebral injury(Cramer’s V=0.817,P<0.001),indicating that the injury was mainly distributed at the violent attack site.There was a moderate correlation between the stress site and the injury site in decelerated craniocerebral injury(Cramer’s V=0.496,P<0.001),indicating that the injury was not only distributed at the violent impact site but also at other sites except the violent impact site.There were significant differences in stress site and injury site between accelerated and decelerated craniocerebral injury(P<0.05).The stress site and injury site of accelerated craniocerebral injury were the most common in the temporal region,followed by the frontal region.The most common stress site of decelerated craniocerebral injury was temporal region,followed by occipital region.The most common injury site was the frontal region and followed by the temporal region in the decelerated craniocerebral injury.There were significant differences in stress site and injury site between unarmed injury and tool injury(P<0.05).The most common injury site and stress site was temporal region and followed by frontal region in the unarmed injury.The most common injury site and stress site of tool injury was temporal region,followed by the parietal region.There was no significant difference in the stress and injury distribution between high fall and flat fall,high fall and traffic accident,and flat fall and traffic accident(P>0.05).There were significant differences in the distribution of craniocerebral injury types between accelerated and decelerated craniocerebral injury(P<0.05).Scalp injury(including scalp abrasion,scalp contusion and laceration,scalp hematoma),skull fracture and epidural hematoma were more common in accelerated craniocerebral injury.Subdural hematoma,subarachnoid hemorrhage,cerebral fracture and ventricular hemorrhage were more common in decelerated craniocerebral injury.2.In the traning dataset,the Dice coefficient of U-Net semantic segmentation model was up to 0.76.In the testing dataset,the Dice coefficient of U-Net semantic segmentation model was up to 0.73.3.The accuracy of Vgg16 model was 89%,and the precision,recall,F1-value and AUC were 94%,97%,95%and 0.98 respectively in the recognition of accelerated craniocerebral injury images;the precision,recall,F1-value and AUC were 75%,67%,71%and 0.88 respectively in the recognition of decelerated craniocerebral injury image;the precision,recall,F1-value and AUC were 90%,91%,91%and 0.90 respectively in the recognition of normal brain images.The accuracy of Res Net18 model was 90%,and the precision,recall,F1-value and AUC were 91%,97%,94%and 0.99 respectively in the recognition of accelerated craniocerebral injury images;the precision,recall,F1-value and AUC were 81%,72%,76%and0.98 respectively in the recognition of decelerated craniocerebral injury images;the precision,recall,F1-value and AUC of normal brain images were 91%,91%,91%and 0.95 respectively in the recognition of normal brain images.The accuracy of Inception＿v3 model was 87%,and the precision,recall,F1-value and AUC were 84%,90%,87%and 0.98 respectively in the recognition of accelerated craniocerebral injury images;the precision,recall,F1-value and AUC were 87%,72%,79%and0.92 respectively in recognition of decelerated craniocerebral injury images;the precision,recall,F1-value and AUC were 89%,90%,89%and 0.93 respectively in recognition of normal brain images.By comparison,it is found that Res Net18 model has many indexes higher than Vgg16 model and Inception＿v3 model,and its performance is the best in this three classification study.ConclusionCT technology could objectively reflect the morphological characteristics and the distribution of craniocerebral injury,and provided a basis for identifying accelerated injury and decelerated injury and inferring the injury manner.Deep learning could automatically segment the scalp hematoma region and the intracranial hemorrhage region and classify accelerated injury and decelerated injury according to craniocerebral CT images,which confirmed the feasibility of deep learning in assisting in imaging identification of accelerated craniocerebral injury and decelerated craniocerebral injury.The above two methods improved the theory and practice of the identification of accelerated craniocerebral injury and decelerated craniocerebral injury from different angles,and could help to enhance the scientificity and objectivity of expert conclusion.