Fully automated radiological analysis of spinal disorders and deformities: a deep learning approach

European spine journal

article

Fabio Galbusera¹  Frank Niemeyer²  Hans-Joachim Wilke²  Tito Bassani¹  Gloria Casaroli¹  Carla Anania³  Francesco Costa³  Marco Brayda-Bruno⁴  Luca Maria Sconfienza⁵ 

[1] Laboratory of Biological Structures Mechanics, IRCCS Istituto Ortopedico Galeazzi;Institute of Orthopedic Research and Biomechanics, Center for Trauma Research Ulm, Ulm University;Department of Neurosurgery, Humanitas Research Hospital;Department of Spine Surgery III, IRCCS Istituto Ortopedico Galeazzi;Unit of Diagnostic and Interventional Radiology, IRCCS Istituto Ortopedico Galeazzi;Department of Biomedical Sciences for Health, Università degli Studi di Milano

关键词: Deep learning;    Spine deformities;    Automated analysis;    Coordinate regression;    Biplanar radiographs;   

DOI  :  10.1007/s00586-019-05944-z

来源: Springer

【 摘 要 】

We present an automated method for extracting anatomical parameters from biplanar radiographs of the spine, which is able to deal with a wide scenario of conditions, including sagittal and coronal deformities, degenerative phenomena as well as images acquired with different fields of view. The location of 78 landmarks (end plate centers, hip joint centers, and margins of the S1 end plate) was extracted from three-dimensional reconstructions of 493 spines of patients suffering from various disorders, including adolescent idiopathic scoliosis, adult deformities, and spinal stenosis. A fully convolutional neural network featuring an additional differentiable spatial to numerical (DSNT) layer was trained to predict the location of each landmark. The values of some parameters (T4–T12 kyphosis, L1–L5 lordosis, Cobb angle of scoliosis, pelvic incidence, sacral slope, and pelvic tilt) were then calculated based on the landmarks’ locations. A quantitative comparison between the predicted parameters and the ground truth was performed on a set of 50 patients. The spine shape predicted by the models was perceptually convincing in all cases. All predicted parameters were strongly correlated with the ground truth. However, the standard errors of the estimated parameters ranged from 2.7° (for the pelvic tilt) to 11.5° (for the L1–L5 lordosis). The proposed method is able to automatically determine the spine shape in biplanar radiographs and calculate anatomical and posture parameters in a wide scenario of clinical conditions with a very good visual performance, despite limitations highlighted by the statistical analysis of the results. These slides can be retrieved under Electronic Supplementary Material.

【 授权许可】

Unknown