Application of computer vision for localization of vertebrae on midsagittal computed tomography slices

A. R. Teplyakova, R. V. Shershnev

Obninsk Institute for Nuclear Power Engineering of the National Research Nuclear University MEPhI

Abstract: utomation of routine operations related to medical image analysis is an important task, as it reduces the workload of radiologists. The selection of computed tomography images corresponding to the levels of specific vertebrae for assessing the patient's body composition is usually done manually, which requires additional time. The purpose is to develop an approach to solving the problem of vertebrae localization on midsagittal computed tomography slices for automatic selection of axial slices used to assess body composition. We developed an approach based on the use of a multiclass segmentation model with the U-Net family architecture and computer vision methods for images preprocessing and segmentation masks postprocessing. In order to assess the impact of input data types and model architectures on segmentation accuracy, we considered 20 approach configurations. We found that the proposed method of preprocessing input data, based on the formation of three-channel images, increases the accuracy of multiclass segmentation for four architectures out of five considered (Dense U-Net demonstrates the maximum Dice similarity coefficient of 0.8858). We also found that the proposed training set augmentation method based on skipping axial slices when forming sagittal slices improves the multiclass segmentation accuracy for models with the ResU-Net and Dense U-Net architectures. Based on the proposed approach, we implemented a software module that solves the problems of automatic determination of the positions of the cervical, thoracic and lumbar vertebrae on the midsagittal computed tomography slice, their visualization and determination of the axial slice indices corresponding to the vertebral body centers. We integrated the developed module with the program for visualization and analysis of DICOM medical files. The developed module can be used as an auxiliary tool in solving diagnostic problems.

Keywords: computer vision, deep learning, computed tomography, segmentation, localization, preprocessing, augmentation, vertebra, U-Net, medical images

DOI: 10.15514/ISPRAS-2025-37(4)-14