Research And Application Of Clavicle Fracture Detection Based On Deep Learning

Clavicle fracture is a common shoulder injury.Clinically,it can be divided into three types of distal,middle and proximal fractures according to the Allman classification method.Doctors make corresponding treatment plans for different fracture types,and different fracture types have different healing standards,healing time and healing degree.Doctors may misdiagnose clavicle fracture due to fuzzy fracture line and other reasons.With the development of artificial intelligence,fracture detection has been increasingly applied in clinical practice,providing doctors with auxiliary diagnosis and improving diagnostic efficiency and accuracy.Fracture detection is currently based on X-ray or CT images.Most studies focus on categorizing section data directly or 3D reconstruction of CT sections and then categorizing.CT image is a body image with a certain thickness in a certain part.If it is converted into slice data for feature extraction or reconstruction,information will be lost.However,the effect of directly classifying 3D CT images is limited by the size of data set and the number of parameters is huge.In order to solve the above problems,a two-stage clavicle fracture detection method was proposed in this paper to realize the detection process of 3D segmentation to extract fracture area,and then screen out key sections for information fusion of CT coronal sagittal plane and cross section,so as to classify clavicle fracture and provide auxiliary diagnosis for doctors.The main research contents and achievements are as follows:First of all,in view of the lack of available public data set for the study of clavicular fracture,CT of normal clavicular fracture and clavicular fracture were collected from the hospital for the study.After preprocessing the collected data,a standardized data set that can be used for network training was obtained.Secondly,aiming at the problems of fuzzy fracture images and unclear fracture line diagnosis difficulties,this paper proposes an improved segmentation algorithm based on 3D UNet as the first stage of the two-stage clavicle fracture detection method.Residual structure,dense connection and inception are combined with 3D U-Net to segment the fracture area.The 3D image after segmentation can provide doctors with a clearer fracture display.The experiment shows that the improved segmentation algorithm proposed in this paper has a better segmentation accuracy and can be more accurate when processing some complex fracture images.Then,this thesis proposes a key layer selection method based on multidimensional entropy to achieve a more accurate selection of slice data for the second stage classification.Then,in order to make better use of the three-dimensional information of the image,this paper proposed a classification method based on multi-view fusion,and carried out feature fusion on the key layer sections from the three perspectives of coronal plane sagittal plane and cross section.Experiments showed that the classification algorithm based on multi-view fusion had better classification effect compared with the section data from a single perspective.Finally,heat map was used to label the fracture area.Improve the interpretability of the model.Finally,the clavicle fracture diagnosis system is designed and implemented.The system is oriented to the doctor group.The doctor introduces the shoulder joint CT of the patient,and then can resamsampling,normalized pretreatment operation,and then can extract the fracture area of the patient CT on the main page,display the segmented fracture area image,and then can conduct classification prediction,display the fracture type.Finally,the diagnosis report is generated and treatment suggestions are given to improve the efficiency of doctors’ diagnosis.