Feasibility Study Of Analyzing Hip Joint-related Parameters In A Population Based On An Intelligent Measurement Model

Purpose: To analyze hip-related parameters based on a deep learning(DL)measurement model and to establish personalized reference intervals(pr RIs)for the discrimination of adult hip dysplasia.Methods: In our previous study,we developed a DL model of hip joints that automatically measures the center-edge(CE)angle,Sharp angle,T(?)nnis angle,and femoral head extrusion index(FHEI)based on anteroposterior(AP)pelvic radiography.Patients(≥17 years of age)who underwent AP pelvic radiographs examination were selected at our hospital(Institution 1)and other hospital(Institution 2)between January2018 and December 2020.A total of 6191 cases were included for our study.397 radiographs from two institutions were used as internal and external data,respectively,to validate the model accuracy and generalization performance.Three radiologists manually labeled the data separately and used their labels to obtain the mean of each parameter as the reference standard.The deep learning model was compared with the reference standard using the Percentage of Critical Points(PCK),intraclass correlation coefficient(ICC),Pearson correlation coefficient(r).Secondly,5,734 cases from Institution 1 were used to analyze the epidemiological trends in hip joint parameters and establish the pr RIs.Linear regression coefficients(β)were calculated to assess the trends in hip parameters according to age and sex,and the values of the mean ± twice the standard deviation of the parameters measured by the model were used as pr RIs.Finally,Radiographs from 30 cases of adult hip dysplasia and paired healthy controls were used to compare the accuracy,sensitivity,and specificity of pr RIs with those of population-based reference intervals(pop RIs).Results:(1)The PCK at the 3-mm distance threshold ranged from 83%–96% and87%–99% in cases from institutions 1 and 2,respectively.There were good correlation and consistency of the four parameters between the model and the reference standard(Institution 1:ICC=0.79-0.88,r=0.81-0.91;Institution 2:ICC=0.71-0.97,r=0.83-0.96).(2)The CE angle tended to increase with age(β,0.16–0.32;P<0.001),whereas the Sharp angle and FHEI decreased with age(β,-0.14 to-0.33;P<0.001).The T(?)nnis angle showed no significant correlation with age in either the left or right hip of males(β,0.013 and 0.009,respectively;all P>0.05).(3)The established pr RIs were able to better identify adult hip dysplasia than the pop RIs(accuracy,85.0% vs 75.0%;sensitivity,92.0% vs 82.6%;specificity,80.0% vs70.3%).Conclusions: The DL model has a good performance in landmark recognition and parameter measurement.The model can accurately obtain hip measurement parameters and establish pr RIs.Compared with pop RIs,pr RIs has good application value in the differentiation of adult hip dysplasia.