ESPE Abstracts

Objective: To investigate the feasibility of constructing a linear regression model to predict the remaining growth potential of adolescent boys based on the radiomics features of knee DR images and clinical characteristics.

Methods: Twenty-four adolescent boys with normal growth and whose knee epiphyses had not yet closed were included in the study of anterior posterio DR images of the knee and the height at the time of filming, and approximate adult height followed up by telephone. The regions of interest (ROI) of the epiphysis and metaphysis of distal femur and proximal tibia were outlined using labelme software, and the radiomics features were extracted using the pyradiomics package on the python. Analysis of variance (ANOVA) and lasso algorithm (LASSO) were used to screen the radiomics features. The radiomics features were subjected to variance inflation factor (VIF) test to remove highly correlated independent variables. The radiomics features that passed the VIF test were combined with height at the time of radiographs to create linear regression equations predicting residual growth potential using the ordinary least squares (OLS) method.

Results: The number of radiomics features was 474, extracted from the femoral and tibial ROIs respectively. ANOVA and LASSO were used to screen the radiomics features. 6 radiomics features were screened for femoral ROI and 7 radiomics features were screened for tibial ROI. 2 radiomics features of distal femur and 3 radiomics features of proximal tibia were excluded by VIF test. In total, there were 8 radiomics features features and the height at the time of filming. Residual growth potential was calculated using height at the time of radiographs and approximate adult height from follow-up. Linear regression equations for predicting residual growth potential were developed using the OLS method. The R2 value of the linear regression equation was 0.86, and the adjusted R2 value was 0.78, which could explain the variation of residual growth potential well, and was statistically significant (P <0.01). In the equation,1 proximal tibia radiomics feature and height at the time of radiographs had a significant effect on residual growth potential (P <0.05).

Conclusion: A linear regression model based on knee radiomics features and height at the time of filming can predict the remaining growth potential of adolescent boys well. Radiomics features of the proximal tibial had a significant effect on the prediction of residual growth potential in adolescent boys.