Differentiating age and sex in vertebral body CT scans – Texture analysis versus deep learning approach

Abstract

The automated analysis of computed tomography (CT) scans of vertebrae, for the purpose of determining an individual's age and sex constitutes a vital area of research. Accurate assessment of bone age in children facilitates the monitoring of their growth and development. Moreover, the determination of both age and sex has significant relevance in various legal contexts involving human remains. We have built a dataset comprising CT scans of vertebral bodies from 166 patients of diverse genders, acquired during routine cardiac examinations. These images were rescaled to 8-bit data, and textural features were computed using the qMaZda software. The results were analysed employing conventional machine learning techniques and deep convolutional networks. The regression model, developed for the automatic estimation of bone age, accurately determined patients' ages, with a mean absolute error of 3.14 years and R2 = 0.79. In the context of classifying patient gender through textural analysis supported by machine learning, we achieved an accuracy of 69 %. However, the application of deep convolutional networks for this task yielded a slightly lower accuracy of 59 %.