Study On Computer-Aided Diagnosis Based On Image Of Distal Radius Fracture

Purpose: With the continuous development of computer recognition systems for processing medical images,object detection algorithms based on deep learning have gradually become well-known and widely applied in various fields,including natural image detection,facial recognition technology,and intelligent driving,all of which have achieved good results.At the same time,the simple and efficient detection process quickly led to new developments in object detection algorithms in the field of medical imaging,such as lung nodule detection,femoral head fracture detection,and lesion area detection of other common diseases.However,in the detection task of distal radius fracture images,there are still some difficulties that need to be solved.The identification and localization of distal radius fracture images in clinical practice(based on the severity of the distal radius fracture and the location of the lesion area as the basis for classification and localization)is an important step in fracture treatment.However,the identification of distal radius fracture images is a tedious and time-consuming task,especially for younger physicians,the lack of clinical experience can lead to a sudden increase in the difficulty of their work,At the same time,unstable situations such as missed inspections caused by work fatigue may also occur.Therefore,how to quickly and accurately classify and locate images of distal radius fractures in clinical diagnosis has become a key issue.The focus of this study is on improving detection efficiency and reducing misdiagnosis rates caused by irregular lesions.Methods: The experimental data used in this article was collected by members of the research group and combined with a publicly available dataset.Based on the imaging features of distal radius images,two experiments were conducted: Experiment 1 was based on the YOLOv4 detection framework.Firstly,improvements were made to the feature extraction network by adding enhanced feature modules such as CSP.The comparison results showed that the network lacked sensitivity to fine lesions in recognizing distal radius images,Therefore,improvements and comparative experiments were made by integrating three attention mechanisms in the three positions of YOLO-v4.At the same time,based on the difficulty of X-ray recognition,several classic convolutional neural network models and object detection network models were compared.Finally,YOLO-v4 based on the fused attention mechanism performed better on the distal radius image dataset.However,in the pre experimental comparison,it was found that YOLOv4 had a slow detection rate in the images of distal radius fractures,and the accuracy of detecting more blurry lesions was unstable.Therefore,in the second part of the experiment,a simpler SSD network model was used,which can effectively improve timeliness and is more suitable for clinical detection requirements.On the basis of single stage target detection,experiment 2 adopted the SSD network model with high detection efficiency according to the imaging characteristics of the distal radius fracture image.At the same time,in order to increase the Receptive field of the network,expand the channel parallel of the model,and integrate the RFB module into the SSD network model.The experimental results obtained by the RFB-SSD network model are better than the data from Experiment 1,with higher experimental speed and accuracy.It can be found that RFB-SDD is more suitable for detecting small target images such as fracture lesion areas.This article is the first to use a single stage object detection YOLO v4 and SSD network model in the distal radius image,showcasing the characteristics of fast operation speed,high accuracy,and concise network model in single stage object detection.Through experiments 1 and 2,it was found that although the detection rate of lesions is similar to that of multi-stage object detection networks in the same period,the detection rate has obvious advantages and is more suitable for clinical fracture detection,Helps physicians preliminarily classify and detect fracture categories.Results: This article constructs two detection and classification models based on images of distal radius fractures,both of which have achieved good expected results.The accuracy of detection in the YOLOv4 algorithm incorporating attention mechanism is 80.40%,while the accuracy in the SSD object detection algorithm is 84.60%.The accuracy of the SSD network incorporating RFB module is 85.83%.Through data comparison,both methods have higher accuracy and detection advantages in classic fracture diagnosis models.Conclusion: In this study,the single-stage target detection algorithm has achieved good detection results in the detection and classification of clinical distal radius fracture images.At the same time,the single-target detection algorithm has the characteristics of fast operation efficiency and high detection speed,which is of certain significance for the application to the field of assistant clinical physician detection.Future research also needs to improve the accuracy of the model,increase the detection rate of the model for small target areas such as small fracture lesions,and further optimize the experimental design.