Using the Mask-RCNN convolutional neural network to automate the construction of two-dimensional solid vertebral models

Biomechanical modeling requires the construction of an accurate solid model of the object under study based on the data of a particular patient. This problem can be solved manually using modern software packages for medical data processing or using computer-aided design systems. This approach is used by many researchers and allows you to create accurate solid models, but is time consuming. In this regard, the automation of the construction of solid models suitable for performing biomechanical calculations is an urgent task and can be carried out using neural network technologies. This study presents the implementation of one of the methods for processing computed tomography data in order to create two-dimensional accurate solid models of vertebral bodies in a sagittal projection. An artificial neural network Mask-RCNN was used for automatic recognition of vertebrae. The assessment of the quality of the automatic recognition performed by the neural network was carried out on the basis of comparison with the Sörensen measure with manual segmentation performed by practitioners. Application of the method makes it possible to significantly speed up the process of modeling bone structures of the spine in 2D mode. The implemented technique was used in the development of a solid-state model module, which is included in the SmartPlan Ortho 2D medical decision support system developed at Saratov State University within the framework of the Advanced Research Foundation project.