| Background: In recent years,the development of information technology,cloud computing,and the improvement of ubiquitous perceptual data and graphics processors have promoted the development of artificial intelligence technology.As a result,the "technical gap" between many sciences and applications has been crossed.Especially in image classification,artificial intelligence algorithms and their applications are gradually maturing.In the hospital,the medical images are virtual proof of clinical diagnosis.Identifying medical images by artificial intelligence is also one of the critical areas of artificial intelligence research.In orthopedics and trauma,discriminating and making classification of fracture on CT and X-ray by artificial intelligence can reduce the workload of doctors and reduce the rate of misdiagnosis and improve the level of diagnosis.In addition,there are many wounded in a short time in natural disasters such as earthquakes and war environments.Artificial intelligence diagnosis can accelerate the diagnosis and diversion of the injured,improve treatment efficiency and save human medical resources.Objective: This study takes calcaneal fractures,which have more complex fracture morphology than other fractures,as the research object.This study aims to make an artificial intelligence model for diagnosing the calcaneal fracture based on CT and X-ray photography with Radiomics and a machine learning algorithm.At the same time,we intended to explore the characteristics of the morphology of calcaneal fractures.Methods: In the first part of this study,we construct the calcaneal fracture 3D model,then draw the calcaneal fracture line map and the calcaneal fracture distribution heat map on the six facets of the calcaneus.This method visually presented the distribution characteristics of the calcaneal fracture line.Then we analyzed the relationship between the distribution of the calcaneal fracture line and the internal structure of calcaneus combined with the distribution of trabeculae in the calcaneus.This part of the research provided an experimental basis for subsequent experiments.In the second part of the study,firstly,we digitized the CT images of fracture and non-fracture calcaneus photography and marked the calcaneal region.Then,we extracted image Radiomics features by Python tool kit and then screened the characteristic variables by the traditional statistical methods combined with machine learning methods.Based on the optimized characteristic variables,an intelligent model for recognizing calcaneal fracture was constructed,and then the model’s accuracy was evaluated.In the third part of the study,two experts classified according to the Sanders method and labeled the calcaneal fractures.Then,an intelligent fracture classification model was constructed through Radiomics and the machine learning method.Finally,the accuracy of the model was evaluated.In the fourth part,based on the X-ray film o,we use the Res Net-18 algorithm of a deep learning network to build an intelligent recognition model for the calcaneal fracture with and without calcaneal markers and evaluate the accuracy of the two final models.Results:(1)We demonstrated the distribution characteristics of calcaneal fracture lines by drawing the fracture line gray maps and heat maps on the six facets of calcaneus.Then,we discussed the relationship between the calcaneal fracture line distribution and the internal structure of calcaneus.(2)We constructed an intelligent recognition model for calcaneal fracture using a Radiomics combined machine learning algorithm.We constructed the model by screening and optimizing characteristic variables and finally built the model based on only72 distinct variables.The overall prediction accuracy was 96% on the test set,the positive recall value(sensitivity)was 92%,and the negative recall value(specificity)was 100%.The accuracy of negative samples was 91%,and the accuracy of positive samples was 100%.The positive f1-score was 96%,and the negative f1-score was 95%.We then simplified the model by screening and optimizing characteristic variables and finally built the model based on only six distinct variables.The overall prediction accuracy was 94% on the test set,the positive recall value(sensitivity)was 92%,and the negative recall value(specificity)was 97%.The accuracy of negative samples was 91%,and the accuracy of positive samples was 97%.The positive f1-score was 95%,and the negative f1-score was 94%.(3)We also constructed an intelligence model by Radiomics and machine learning to classify calcaneal fracture according to Sanders’ s classification method.The overall prediction accuracy of the model was 53% on the test set.For each subtype,the f1-score of type I was 0.45 and the area under the ROC curve was 0.66;Type II f1-score was 0.57,and the area under the ROC curve was0.86;Type III f1-score was 0.46,the area under the ROC curve was 0.63,type IV f1-score was 0.57,and the area under ROC curve was 0.83.To enable the computer to recognize calcaneal fractures on X-rays,we use the depth network algorithm--resnet-18 to build an intelligent prediction model.At last,the accuracy of the resnet-18 model was stable at 95.5%,and f1-score was steady at 0.953 after 15 iterations.Based on the unmarked X-ray film,the accuracy of the resnet-18 prediction model is stable at 85% to 87%,and the f1-score is stable at 0.833 to 0.845 after 12 iterations.After L2 regularization was added to the model,the accuracy was improved to 91%.Conclusion:(1)Although there are significant differences in fracture location,degree of comminution,and displacement of fracture block in different types of calcaneal fractures,we also found some regularities.These regularities are determined by the internal structure of the calcaneus,the position of the calcaneus and talus during fracture,and also determined by the direction and duration of violence.These regularities can evaluate some of the Radiomics features such as texture,gray value changes,2D、3D shape features,etc.Know the changes of Radiomics features caused by the fracture.Mastering this rule is very beneficial to establish an intelligent diagnosis model of a calcaneal fracture.(2).Radiomics,combined with machine learning algorithms such as lasso regression random forest,could automatically discriminate the calcaneal fracture on CT.It has high accuracy and can be used in clinics to diagnose calcaneal fractures rapidly.(3)The intelligent classification model of calcaneal fracture constructed by Radiomics combined with multi-classification machine learning algorithms such as One Vsrest Classifier had an ordinary accuracy.The main reason was that the consistency of Sanders’ s classification was low.It was necessary to improve the model’s accuracy by improving the consistency of Sanders classification or establishing another classification method with high character and repeatability dedicated to machine classification.(4)The resnet-18 algorithm of a deep convolution neural network was used to establish the recognition model of a calcaneal fracture.When the X-ray film has marked,the model had a high recognition accuracy,and the accuracy decreased when the X-ray was not marked.Therefore,using a deep learning algorithm to establish a recognition model based on X-ray film is feasible and highly accurate.However,further optimization of the model and process is still needed to meet clinical application requirements. |
PDF Full Text Request