Research On Intelligent Diagnosis And Image Segmentation Of Osteonecrosis Of The Femoral Head Based On Deep Learning

Osteonecrosis of the femoral head(ONFH)is a potentially devastating disease that is an important cause of hip disability in the young and middle-aged population.The prognosis of ONFH depends on the timing of diagnosis and intervention,the stage and classification of the disease,and the location and size of the necrotic lesion.Once the femoral head collapses or progresses to the middle and late stage,total hip replacement can only be performed.Conversely,if the corresponding hip-preserving surgery is performed in the early stage or before the femoral head collapses,it will help to delay the progression of osteonecrosis,and even reduce the possibility of total hip arthroplasty as the final replacement.Magnetic resonance imaging(MRI)is the most sensitive and specific examination to identify ONFH,and has important application value in the early stage and clinical staging of ONFH.The detection of early necrotic lesions,the classification of disease and the quantitative evaluation of lesions based on MRIs are the key steps and effective means for early diagnosis of ONFH,as well as an important basis for hip-preserving surgery.In recent years,artificial intelligence has been widely used in the field of medical imaging and has shown great potential.In the present study,we explore the use of deep learning technology to study the intelligent detection,classification and automatic segmentation algorithms of ONFH based on MRI,in order to achieve early diagnosis,accurate classification and precise quantification of lesion volume,which can provide individualized surgical decision-making and theoretical basis for hip-preserving surgery.Early hip-preserving surgery for ONFH usually involves three steps: First,timely diagnosis of early ONFH,then grading according to the severity of the disease,predicting the risk of collapse of the femoral head.And finally quantitative evaluation of necrotic lesions was performed based on the MRI of the affected hip to determine the planning path of hip-preserving surgery and bone grafting requirements.The main research content of this question are as follows:Early identification of ONFH is difficult,and the accuracy of diagnosis depends on the clinical experience of orthopaedic surgeons and radiologists.To address this problem,we developed and validated the convolutional neural network(CNN)-Res Ne St model,which can perform preliminary evaluation and lesion detection on MRI.Then we compared the diagnostic performance of the Res Ne St model with orthopaedic surgeons of different experience levels to assess clinical usability.The results of the study showed that the area under the receiver operating characteristic curve(AUC),accuracy,sensitivity and specificity values of the developed Res Ne St model in the diagnosis of early ONFH were 0.98,98.4%,97.6%and 98.6%,respectively.Its diagnostic performance was significantly better than that of orthopaedic attending surgeons and residents,and it can effectively assist junior surgeons to improve the diagnostic accuracy of early ONFH.Due to the irregular shape of femoral head necrosis lesions,it is often difficult to reach a consensus on the classification of the disease,and it depends heavily on the clinical experience of orthopaedic surgeons.Therefore,we proposed a multi-class CNN and optimized it,which can perform corresponding classification according to the location of necrosis lesions.In this chapter,we constructed a multi-center dataset containing ONFH hip MRI from four hospitals in China to solve the problem of small training samples in multi-classification tasks,and adopt the oversampling method to solve the problem of multi-class sample imbalance.To visualize the decision-making attention of the CNN model,the gradient-weighted class activation mapping(Grad-CAM)technique is used to demonstrate that the developed CNN model makes decisions based on the features of necrotic lesions.In addition,we conducted a comparative study of the developed CNN model with different levels of orthopaedic surgeons,and conducted rigorous multi-center external validation to evaluate the generalization of the CNN model in real clinical settings.In internal validation,the overall accuracy of the CNN model for predicting the severity of ONFH based on Japanese Investigation Committee(JIC)classification was 87.8%.The macro-average values of AUC,precision,recall,and F-value were 0.90,84.8%,84.8%,and 84.6%,respectively.In external validation,the overall accuracy of the CNN model was83.8%.The macro-average values of AUC,precision,recall,and F-value were 0.87,79.5%,80.5%,and 79.9%,respectively.In a further human-machine comparison study,the results showed that the classification accuracy of the developed multi-class CNN model was outperformed or comparable to that of the deputy chief orthopaedic surgeons,but there was no statistical difference.The Grad-CAM heatmap showed that the decision-making of the developed multi-class CNN model was mainly activated by the necrotic region of interest,confirming its reliable performance for performing JIC classification.In order to efficiently segment the MRI-based hip joint bone model and necrotic lesions of different sizes and shapes,in this chapter,we propose an intelligent segmentation model of Res-U net.The CNN model can automatically segment the bone structures of the hip joint by inputting MRI images,and determine the characteristics of necrotic lesions at different levels and scales.Meanwhile,on the basis of accurate segmentation,we calculated the volume of necrotic lesions by stacking two-dimensional segmentation masks,which can reliably achieve three-dimensional quantitative evaluation of ONFH necrotic lesion volume.The results show that the proposed CNN segmentation model achieves 93.4%,90.9%,and78.0% dice similarity coefficient for acetabulum,unaffected femoral head,and necrotic lesions,respectively.The intraclass correlation coefficient of the volume of necrotic lesions measured between manual segmentation and CNN segmentation was0.82,indicating that the two methods have good agreement.The results of three-dimensional quantification of necrotic lesions showed that the average volume of necrotic lesions in the included patients was 9.70 ± 8.44cm3.The intelligent diagnosis,classification and segmentation model based on deep learning technology has shown great advantages in the clinical application of early ONFH.It can be used to provide fast and accurate diagnosis and classification in clinical practice,and can assist orthopaedic surgeons in the preoperative planning of early ONFH hip-preserving surgery.