Research On Target Detection Algorithm And Detection Network Structure Of X-ray Fracture Line Based On Convolution Neural Network

Part One Study on the structure and detection performance of a two-stage target detection network for fracture lines in X-ray plain films based on convolutional neural networksObjective: A two-stage object detection network Faster R-CNN was built to detect fracture lines on X-ray films,and a multi-scale feature pyramid networks(FPN)and region of interest align(ROI Align)a module were added to solve the problem of small target detection and candidate box regression position deviation on the large feature map,and realize the accurate detection and localization of small targets such as fine fracture lines and multiple fracture lines.Methods: A total of 7984 X-ray images of acute and subacute traumatic fractures were collected,and the fracture data set was randomly divided into training set,validation set,and test set at a ratio of 7:2:1.After image preprocessing,the two-stage target detection network built by mmdetection toolbox was input to detect the fracture line.Faster R-CNN was set as the baseline model,and Faster R-CNN+FPN+ROI Align was used as the experimental model to perform ablation experiments on FPN,ROI Align and other modules.Using precision,sensitivity/recall,F1-Score,average precision(AP),accuracy,specificity and other indicators were used to evaluate the network to verify the research hypothesis.Results: The AP value increased by 2.9% after adding the ROI Align module,the AP value was increased by 6.5% after adding the FPN module,and the experimental model increased the AP value by 9.4%.The addition of FPN to the baseline model improved the detection of small avulsion fracture,nondisplaced linear fracture line,slight trabecular bone distortion,and cortical bone fold.The accuracy rate,recall rate,F1 value,accuracy and specificity were increased by 3.6%,5.3%,4.7%,6.5% and 2.2%,respectively,compared with the baseline model.The addition of ROI Align module improves the detection of multiple fractures of the hand and foot.The accuracy rate,recall rate,F1 value,accuracy and specificity were increased by 2.3%、4.0%、3.4%、2.6%、and1.2%,respectively,compared with the baseline model.Compared with the baseline model,The precision,recall,F1 score,accuracy and specificity of the experimental model were increased by 5.7% and 8.6%,7.5%,6.1%,3.7%respectively.Conclusions: The FPN module and ROI Align module enhance the precise detection and localization of small and multiple fractures,which verified the hypothesis that the Faster R-CNN+FPN+ROI Align deep learning model could effectively detect and localize fracture lines in X-ray plain films more accurately.The results show the potential of CNN-based object detection methods for plain X-ray fracture diagnosis.Part Two Algorithm research on reducing feature ambiguity of fracture lines in target detection network and efficiency analysis of the feature ambiguity mitigate operator model applied to X-ray plain film fracture diagnosisObjective: This study was performed to propose a method,the feature ambiguity mitigate operator(FAMO)model,to mitigate feature ambiguity in bone fracture detection on radiographs of various body parts.Methods: According to the morphology and spatial relationship of the fracture line,FAMO model was developed to act on the second stage of the twostage target detection method.The fracture lines in the images were segmented and labeled by three radiologists.The initial encoder structure was constructed as the Residual Network(Res Ne Xt101)+ feature pyramid networks(FPN)for object detection.The preprocessed plain X-ray films were input into Res Ne Xt101,and the feature maps of different proportions generated by Res Ne Xt101 were collected by FPN,and then the feature maps were ambiguous reduced by FAMO.The Res Ne Xt101+FPN+ region of interest align(ROI Align)module was used as the control model,and the FAMO model was used as the experimental model to test on the test set.To the per-fracture extent,an AP(average precision)analysis was performed.For per-image and per-case,the sensitivity,specificity and area under the receiver operating characteristic curve(AUC)were analyzed.Results: The diagnostic efficacy evaluation indexes of FAMO model were better than those of control model.Compared with the control model,FAMO increased AP value from 76.8% to 77.4% in fracture line detection.The improvements in sensitivity and specificity per image level for fracture diagnosis were 2.6%(from 61.9% to 64.5%)and 1.4%(from 91.5% to 92.9%),respectively,and the improvements in AUC were 2.6%(from 74.9% to 77.5%).Finally,the sensitivity of FAMO model increased from 74.9% to 77.5%,the specificity increased from 91.7% to 93.4%,and the AUC increased by 0.8%(from 85.7% to 86.5%).In the elbow,shoulder,and knee with a small sample size,the FAMO model at the per-case level was difficult to identify elbow fractures(AUC was 71.1%),but easy to identify shoulder fractures(AUC was83.2%).Among the body parts with a large sample size,FAMO detected pelvic fracture with an AUC of 96.9% at the per case level,while FAMO detected foot fracture with an AUC of 82.1% at the per case level.Conclusions: In the deep learning convolutional neural fracture detection of plain X-ray films,FAMO model can improve the diagnostic performance of the convolutional neural network model by reducing the feature ambiguity of the fracture line in the feature map.