Classification Study On X-Ray Images Of Humeral Shaft Fractures By CBIR Technique Based On Traditional And Deep Features

Research purpose: As a common type of disease in daily life,bone fractures' occurrence will cause serious harm,and if not timely treated,it will cause serious consequences.Humeral shaft fracture is a common type of bone fractures,whose occurrence will not only damage the nerve and blood vessels,but also make the shoulder elbow joint dysfunction.Xray imaging is a common method used in clinical fracture diagnosis,but with the continuous development of computer technology and the increasing dependence of physicians on medical image data,the number of medical images increased a lot,which put physicians under severe stress and more and more work.As one of the Computer-aided Diagnostic techniques,the Content-based Image Retrieval(CBIR)can retrieve similar cases from largescale medical image databases,thus helping doctors to diagnose a given image data type,which plays an important role in the classification of medical images.Therefore,based on the CBIR technique,the X-ray images of humeral shaft fracture were classified into fracture and non-fracture in our paper.Research contents: This paper studies the classification of humeral shaft fracture from two aspect of traditional features and deep features,and mainly did the work as follows: 1.Firstly,the paper classified the humeral fracture images based on the traditional features.In this paper,we first extracted the traditional texture features of humeral shaft fracture for classification analysis,including LBP feature,Gabor feature and Haralick feature,employing the Euclidean distance based on K-Nearest Neighbor algorithm to measure the similarity of features,and evaluated the classification performance from the number of retrieval images(K value)and feature analysis.2.Classification study of humeral shaft fractures based on deep features.Secondly,the deep features were used to classify the humeral shaft fracture images,the Inception-V3 structure of Goog Le Net and Res Net network were used to extract deep features from X-ray images of humeral shaft fractures,and the PCA algorithm was employed to reduce the dimension.The measurement of similarity was also based on the K-Nearest Neighbor algorithm employing the Euclidean distance.The performance of the proposed scheme was evaluated from the number of retrieved images and the classification of features.The AUC mean value and classification accuracy were used as the evaluation criteria for the two experiments.Research results: As for traditional features,the AUC mean value reached highest when the K value was 50 and the AUC mean value and classification accuracy reached the highest when the feature was combined-feature,which were 0.973 and 0.915 respectively.And as for deep features,the AUC mean value reached highest when the K value was 10 and the AUC mean value and classification accuracy reached highest,when the feature was combined-feature,which were 0.985 and 0.949 respectively.Conclusion: In this study,we used the CBIR technology to study the classification of humeral shaft fracture images based on traditional texture features and deep features,and achieved good results.In the following research,we will do further research on data preprocessing,feature selection and the type of images needed for classification to get higher accuracies for a wider classification of bone fracture images.