부산대학교 도서관

in order to create a digital twin of the spine to simulate the behavior of the spine after various interventions, the task of automating the marking of the vertebral bodies on radiographs was formulated. The presented work describes the developed approaches to solving the problem of marking the vertebral bodies on radiographs, and also describes algorithms for extracting images of the spine and images of individual vertebrae. For the detection of images of each individual vertebra, the YOLOv5 model was used, and an ensemble of regression models was created to solve the problem of marking the image of each individual vertebra. After a series of experiments, marking errors were estimated at 4.6% using the ratio of coordinates obtained using trained models and correctly marked data. Theories were also put forward about the possible improvement of the created software by dividing the vertebrae into two subclasses in order to reduce the markup error to the desired 0.8%.