Research On Screening Method Of Scoliosis Based On Convolutional Neural Network

Scoliosis is a common disease especially during the growth of adolescents.In China,the incidence of idiopathic scoliosis in adolescents is about 1%to 3%.However,it is not easy to detect at the early stage of the disease.During the critical period of physical growth and development,teenagers are vulnerable to external influences.Due to the lack of exercise,excessive study and incorrect walking posture,the spine is easily affected.If not treated in time,it might bring irreversible damage to patient's spine and seriously affect the patient's body shape,mental health and even their life.Normally orthopaedic surgeon will manually measure and calculate the Cobb(named after surgeon John Robert Cobb)based on the shape of the spine presented by the patient's radiograph to determine if patients have scoliosis and its severity.This diagnosis method will inevitably expose patients to X-ray radiation,increase the workload of doctors,and waste a lot of medical resources.With the rapid development of computer-based artificial intelligence technology,computer-aided diagnosis has become an important means to assist doctors in diagnosis,and it has also achieved certain results,which provides a new direction for current medical situation of scoliosis.In this thesis,X-rays were discarded,and the back-up images of patients with non-radiation and easy-to-acquire method were selected.The screening method of scoliosis was studied based on convolutional neural network.The study in this thesis contains the location of the region of interest in the patient's back image and the grading diagnosis of scoliosis.In the study of the location of the region of interest in the patient's back image,this thesis firstly compares the R-CNN,Fast R-CNN and Faster R-CNN target detection convolutional neural networks,and finally chooses the Faster R-CNN volume by factors such as accuracy.The neural network locates the back area of the patient.In the grading study of scoliosis,this article firstly combines with the doctor's clinical experience,according to the size of the patient's spinal Cobb angle,divided into unaffected(Cobb angle of 0o~10o),mild scoliosis(Cobb angle of 11o~25o),moderate scoliosis(Cobb angle 26o~45o)and severe scoliosis(Cobb angle>45o)four levels.Then,through a variety of classical convolutional neural networks(AlexNet,VGG,DenseNet and ResNet)for comparison.Taking into account factors such as accuracy and duration of model training,the ResNet convolutional neural network was used to classify the scoliosis disease in detail,and the different parameters and network structure in the ResNet were used to classify the scoliosis.The impact of performance,which in turn optimizes the network.In the end,we conduct an experiment on the convolutional neural network method,the classical feature extraction method(color and texture composite feature,local binary mode)as well as the SVM(Support Vector Machine)in machine learning to find the optimal method.In this thesis,the four-fold cross-validation method is used to increase the reliability and the generalization ability of the model.In the study of the location of the region of interest in the patient's back image,using the Faster R-CNN convolutional neural network,the area size of the network parameter anchor box is(64~2,128~2and 256~2),and the number of proposal is 256.The positioning accuracy reach up to 99.17%.In the grading diagnosis of scoliosis,this thesis first uses the ResNet 50 convolutional neural network for four levels of classification(four classification model),the accuracy of the model is 65.17%,which is relatively low.Combined with the actual screening of scoliosis,this thesis finally uses three binary classifications for detailed grading(a two classification models of scoliosis,a two-class model of mild scoliosis and a two-class model of severe scoliosis).The accuracy rate can reach 91.23%,86.92%and 82.45%respectively.The overall grading effect exceeds the four-class model and the combination of classical feature extraction and SVM classifier.Based on the experimental results,the results of the combination of Faster R-CNN and three ResNet 50 two-category models studied in this thesis can be used as a reference for orthopedic surgeons to diagnose scoliosis.