This paper presents the development of a software module for the automatic, accurate, and reproducible measurement of the Cobb angle-the “gold standard” for assessing scoliotic deformity on anteroposterior spinal radiographs. The solution is based on a hybrid approach combining deep learning for semantic segmentation of vertebral bodies and deterministic computer vision algorithms for subsequent geometric calculation. A key feature of the proposed approach is the utilization of the U2-Net architecture, which demonstrates high efficiency on a heterogeneous dataset (313 images) comprising both DICOM files and photographs of X-ray films on viewing boxes (JPG/PNG). Subsequent algorithmic contour analysis and approximation of vertebral endplates enable the automatic construction of Cobb lines and angle computation. The developed artificial neural network for vertebral body segmentation was validated on a test dataset. The model achieved a Dice coefficient of 0.9859. The module has been integrated into a medical information system for a prospective accuracy evaluation involving expert clinicians.
Download the article (PDF) from the journal's website (in Russian): http://www.isa.ru/proceedings/images/documents/2026-76-3/68-78.pdf
Download the article (PDF) from eLibrary (in Russian, registration required): https://www.elibrary.ru/item.asp?id=93335286
Veits, A., Blagosklonov, N. Automated Cobb Angle Measurement on Spinal Radiographs Using a Hybrid Approach // Proceeding of the Institute for Systems Analysis of the Russian Academy of Science, 2026, volume 76, № 3, pp. 68–78.