| Background:Developmental dysplasia of the hip(DDH)is a common orthopedic disease in children.Patients with early diagnosis of DDH can be effectively treated by conservative means.Delayed diagnosis requires surgical treatment.However,surgical treatment is costly and has a low cure rate,and may have a series of sequelae.Therefore,the early diagnosis of DDH is of great significance.However,the bones of children with DDH are in rapid development and usually show different degrees of deformity,which leads to complex and variable imaging manifestations in key areas of pelvic X-ray and difficult diagnosis.In recent years,with the extensive application of artificial intelligence in the medical field,some scholars proposed a method based on edge detection to extract the accurate acetabular edge from the pelvic X-ray and finally calculate the acetabular index,which can be used as the basis of diagnosis of DDH.Some scholars also propose the method of using template matching to find the obturator foramen in the pelvic X-ray image.This method can determine the remaining key points through the obturator coordinates,thereby calculating the acetabular index and evaluating the condition of patients with DDH.Although these methods have realized the intelligent diagnosis of DDH,they still have some limitations.Under the above background,based on the deep learning method,this study developed an AI system that can diagnose DDH quickly and accurately using a data set containing10,219 standard pelvic plain films of children,and analyzed the feasibility of its clinical application.Objective: 1.To establish a new database based on pelvic imaging data and disease classification of children in the region.To deepen the understanding of the regularity and influencing factors of children with DDH,so as to explore new mechanisms and methods for the diagnosis of children with DDH.2.Explore new ideas and methods for the diagnosis and treatment of DDH patients from the perspective of artificial intelligence,and analyze the feasibility of its clinical application.Finally,it lays a foundation for early diagnosis and accurate intelligent diagnosis of DDH.Methods: 10,219 anonymous pelvic plain films from April 2014 to May 2018 were retrospectively collected: 0-10 years old;0-2 years old: 8577;≥2 years old: 1,642;Male: 2,410;Female: 7,809.All pelvic plain films were marked by multiple rounds of cross-examination according to Hilgenreiner and T?nnis methods.Doctors marked bilateral acetabular center,the lateral edge of the acetabulum point and femoral head epiphysis center through image PACS system.Then the line connecting the center points of bilateral acetabulum is the Hilgenreiner line,and the line perpendicular to the Hilgenreiner line and passing through the lateral edge points of acetabulum is the Perkins line.Finally,the T?nnis grade of pelvic plain film is determined by using the quadrant formed by the intersection of Hilgenreiner line and Perkins line.Patients without epiphysis of femoral head were classified according to the standard of International Hip Dysplasia Institute(IHDI).The angle formed between the line connecting the acetabulum center point and the acetabulum lateral edge point and the Hilgenreiner line is the acetabulum index.The accuracy of pelvic plain film labeling was reviewed by multiple rounds of cross-examination,and the dislocation degree and acetabular index of all children were statistically analyzed.In this study,we proposed a image denoising and quality enhancement method for X-ray image of hip joint in children based on deep convolutional neural network.By automatically extracting the essential features of the original image and image denoising,the accuracy of the system for the detection of x-ray images with different qualities is improved.Firstly,by analyzing the types and distribution characteristics of noises in hip x-ray,a framework of generating noisy images based on generating countermeasure network is designed,and a data set consisting of 2200 noisy images is constructed.Then 2000 noisy images are randomly selected and input into the convolutional neural network to train the network’s denoising and quality enhancement capabilities.The remaining 200 sheets were used as the test set to evaluate the effect of image denoising using Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM).The labeled pelvic plain films were randomly divided into three data sets:(1)8,000pelvic plain films were used to train and build an artificial intelligence aided diagnosis system;(2)1081 pelvic plain films were used to optimize the system;(3)1138 pelvic plain films were used to test the system.The artificial intelligence system is trained with 8000 labeled pelvic plain films,and the key points are located by the method of deep learning coordinate point regression target detection,and finally 12 coordinate values of six key points which are most important for DDH diagnosis are output.In the control experiment,the template matching method was used to identify the key points of pelvic plain film,determine the quadrant position of femoral head and diagnose DDH.Finally,the accuracy of the two experimental methods to identify key points is compared and analyzed.The results of 1138 cases(2276 hips)diagnosed by artificial intelligence and doctors were compared.The consistency between the two methods in diagnosing dislocation of hip joint and T?nnis grading results was evaluated.The consistency of acetabular index measured by the two methods was evaluated.Then 200 samples were randomly selected from the test set and independently measured by another 8 doctors.The consistency of acetabular index measurement results and T?nnis grading diagnosis results of the artificial intelligence system and clinicians was evaluated.Results:1.Total data set of 10219 pelvic X-rays,including 4008 with normal bilateral hip joints(8016 hips);There were 1,496 males,and the mean acetabular index was21.26°±4.26°.There were 2512 females with an average acetabular index of22.51°±3.52°.The difference between the male and female groups was statistically significant(P<0.001).The Spearman correlation coefficient between acetabular index and age was-0.608 in the normal hip group(P<0.001).For 7643 hips with T?nnisⅠdegree dislocation,the Spearman correlation coefficient between acetabular index and age was-0.581,P < 0.001;For 1251 hips with T?nnis II dislocation,the Spearman correlation coefficient between acetabular index and age was-0.237(P < 0.001);887hips of T?nnis III degree dislocation,and the Spearman correlation coefficient between acetabular index and age was 0.068,P = 0.042;For 853 hips with T?nnis IV degree dislocation,the Spearman correlation coefficient between acetabular index and age was0.014,P=0.680.2.After noise reduction,the average PSNR index of 200 test set images reached40.981 d B,which was 31.05% higher than that of the original noisy image of 31.272 d B.In terms of SSIM index,the average image denoising and enhancement algorithm proposed in this paper reached 0.96,which also achieved a significant relative improvement of 47.69% compared with the original noisy image of 0.65.The key point matching error results of deep learning and template matching methods are as follows:Point 1:1.22±0.86 mm,5.82±24.45 mm,P<0.001;Point 2:1.41±0.88 mm,6.85±31.52 mm,P<0.001;Point 3:1.15±0.89 mm,9.22±31.04 mm,P<0.001;Point4:1.24±0.92 mm,13.45±53.75 mm,P<0.001;Point 5:0.91±0.85 mm,11.46±28.74 mm,P<0.001;Point 6:0.95±0.97 mm,9.88±44.00 mm,P<0.001.3.The Kappa consistency coefficient of the overall AI system in T?nnis grading diagnosis compared with clinicians was 0.796,and the accuracy rate was 1998/2276(87.79%).A total of 2276 hips were assessed by the deep learning system and a clinical team to determine the presence or absence of ‘dislocation’.Our analysis of the entire dataset showed that the area under the receiver operating characteristic curve(AUC)for our deep learning system was 0.975,the accuracy was 2254/2276(99%),the sensitivity was 276/289(95.5%),and the specificity was 1978/1987(99.5%).In the group of younger children(1978 hips),the AUC was 0.974,the accuracy was1956/1978(98.9 %),the sensitivity was 265/278(95.3%),and the specificity was1691/1700(99.5%).The infants under 6 months-of-age(190 hips),the AUC was 0.952,the accuracy was 188/190(98.9 %),the sensitivity was 19/21(90.5%),and the specificity was 169/169(100%).In the group of older children(involving 298 hips),the rate of accuracy was 298/298(100%).A comparison of the distribution of acetabular index measurements made by the deep learning system,with those made by clinicians.With regards to acetabular index measurement(2276 hips),there was a 95%limit of agreement,as determined by the Bland-Altman method,of-4.0° to 3.45°(bias=-0.27°,P < 0.0001)compared with measurements made by clinicians.With regards to acetabular index measurements in younger children,the 95% limit of agreement was-4.1° to 3.5°(bias =-0.3°,P < 0.0001),for infants under 6 months-of-age,the 95%limit of agreement was-3.87° to 3.77°(bias =-0.05°,P = 0.728),and for those of older children,the limit was-3.38° to 3.25°(bias =-0.07°,P = 0.5013).The 95% limits of agreement,using the Bland-Altman method,for acetabular index measurement in the confirmed ‘non-dislocation’ and ‘dislocation’ groups were-3.27° to 2.94°(bias =-0.17°,P < 0.0001)and-7.36° to 5.36°(bias =-1.0°,P < 0.0001),respectively.4.Comparison of the measurement evaluations by the system and eight clinicians vs.that of,the expert committee,the 95% LOA of the clinician with the smallest measurement error was-2.76° to 2.56°(bias =-0.10°,P = 0.126).The 95% LOA of the system was-0.93° to 2.86°(bias =-0.03°,P = 0.647).The 95% LOA of the clinician with the largest measurement error was-3.41° to 4.25°(bias = 0.42°,P <0.05).The measurement error of the system was only greater than that of a senior clinician.Conclusions:The acetabular index of hip joint in normal children is negatively correlated with age.The acetabular index of women was slightly higher than that of men and the difference was statistically significant.There is a negative correlation between acetabular index and age in children with T?nnis I and II degree dislocation,but there is no correlation between acetabular index and age in children with T?nnis III and IV degree dislocation.The image denoising and quality enhancement method of deep convolution neural network has remarkable effect,which can effectively remove image noise in hip X-ray images and improve image quality.The deep learning method is more accurate than the template matching method in matching key points of hip X-ray images.The diagnosis result of deep learning system is consistent with that of clinicians,and it is more convenient and faster.The system can be used to diagnose DDH according to pelvic plain film,and improve the complicated screening and referral process of manual diagnosis at present. |
PDF Full Text Request