Research On Classification And Grading Of Traumatic Brain Injury Based On Deep Learning

Among many traumas,traumatic brain injury(TBI)is a serious injury with a very high incidence and the highest rate of death and disability.Identifying the type of trauma through cranial imaging has become an important means of preoperative diagnosis of the brain.Although medical imaging diagnosis has made significant progress,due to the extremely complex and diverse brain imaging in traumatic brain injury,accurate imaging diagnosis is still a difficult task.Deep learning has been proven to be an effective method to improve the performance of medical image analysis.However,due to the lack of a comprehensive traumatic brain injury image data set,current research in this direction is limited.The incidence area of traumatic brain injury is very extensive,and the imaging manifestations are complex and diverse,but the diagnosis time is very urgent,and the medical level of each area is uneven,and some grass-roots areas may lack experienced physicians and advanced equipment,thus missing the rescue critical timing.In response to the above problems and challenges,this work provides a new CT image data set suitable for traumatic brain injury detection,which includes 226(cranial trauma/normal: 175/51)subjects with a total of 6780 slices.Labels are provided by experienced radiologists.With the help of this data set,we first proposed a TBI image diagnosis model based on Convolutional Neural Network(CNN)combined with Recurrent Neural Network(RNN)and Embedded Extrusion and Excitation(SE)module for the classification task of traumatic brain injury.In addition,in order to avoid the problems of local optimization and insufficient data,a transfer learning method is introduced.Experimental results show that the model achieves95.9% accuracy in predicting whether there is damage at the slice level,which is more accurate than other commonly used classification methods.Furthermore,we conducted a grading study of traumatic brain injury.According to the judgment standard provided by the doctor,all patients were divided into three grades: mild,moderate,and severe traumatic brain injury.Further considering that the information obtained by the two-way convolution can be weighted,for the feature extraction of different scales of different lesions,an adaptive and selective convolution kernel is used,and the SK module is used to embed the proposed network.The amount of calculation introduced is small,but Can improve accuracy.Use our model to classify this task to achieve the purpose of classification.The experimental results show that the task accuracy rate for the tri-category traumatic brain injury grading is relatively high,and the project goal is initially achieved.We believe that the current work can help doctors make further clinical diagnosis.