METHODS: A retrospective analysis of prospectively collected data was conducted. Skeletal maturity was determined using Risser, SSMS, TOCI and CVM for each patient. Crosstabulations of axial vs. appendicular markers were formed to analyze their concordance and discordance. Logistic and logarithmic regression models were run to assess longitudinal growth (postoperative height gain and leg-length growth) and curve modulation (follow-up instrumented Cobb correction after index operation), respectively. Models were compared using Akaike information criterion (AIC).
RESULTS: 34 patients (32 F/2 M, mean age: 12.8 ± 1.5 years, mean follow-up: 47.7 (24-80) months) were included. The median preoperative maturity stages were: Risser: 1 (-1-4), SSMS: 4 (1-7), TOCI: 6 (1-8) and CVM: 4 (1-6). At latest follow-up, all patients reached skeletal maturity. Concordance and discordance were observed between axial vs. appendicular systems that demonstrated a range of possible distributions of CVM, where trunk peak height velocity occurred before, simultaneously with or after the standing height peak height velocity. R-squared values for Risser, SSMS, TOCI and CVM were 0.701, 0.783, 0.810 and 0.811, respectively, for prediction of final height; 0.759, 0.821, 0.831 and 0.775 for final leg-length, and 0.507, 0.588, 0.668 and 0.673 for curve modulation. Delta AIC values demonstrated that different skeletal maturity assessment methods provided distinctive information regarding follow-up height gain, leg-length growth and curve behavior.
CONCLUSIONS: Risser score provided considerably less information for all three outcome variables. TOCI and SSMS provided substantial information regarding remaining leg-length assessments, while in terms of assessment of total height gain and curve modulation after surgery, CVM and TOCI offered substantial information and SSMS offered strong information. Mutual use of axial and appendicular markers may provide valuable insight concerning timing of surgery and magnitude of surgical correction.
MATERIAL AND METHODS: The study included 223 tomograms of the head and neck in sagittal projection from patients without any pathology of the studied structures. Morphometric analysis was carried out using PjaPro and Gradient programs, statistical analysis was performed by SPSS Statistics software. A fully convolutional EfficientNet-B2 neural network was used, which was trained in two stages: selection of the area of interest and solution of regression tasks.
RESULTS: Morphometric assessment and subsequent statistical analysis of the selected group of features have shown presence of the strongest correlation with age in the indicator characterizing the involution of the median atlantoaxial joint. A deep learning method using the convolutional network, which automatically selects the desired area in the image (the area of the vertebral junction), classifies the sample, and makes an assumption about the age of the unknown individual with an accuracy of 7.5 to 10.5 years has been tested.
CONCLUSION: As a result of the study, a positive experience has been obtained indicating the possibility of using convolutional neural networks to determine the age of the unknown person, which expands the evidence base and provides new opportunities for determining group-wide personality traits in forensic medicine.