Hubert Labelle

Genetic Algorithm–Neural Network Estimation of Cobb Angle from Torso Asymmetry in Scoliosis

Scoliosis severity, measured by the Cobb angle, was estimated by artificial neural network from indices of torso surface asymmetry using a genetic algorithm to select the optimal set of input torso indices. Estimates of the Cobb angle were accurate within 5 degrees in two-thirds, and within 10 degrees in six-sevenths, of a test set of 115 scans of 48 scoliosis patients, showing promise for future longitudinal studies to detect scoliosis progression without use of X-rays.

Simulation of the postoperative trunk appearance in scoliosis surgery

Persistence of external trunk asymmetry after scoliosis surgical treatment is frequent and difficult to predict by clinicians. This is a significant problem considering that correction of the apparent deformity is a major factor of satisfaction for the patients. A simulation of the correction on the external appearance would allow the clinician to illustrate to the patient the potential result of the surgery and would help in deciding on a surgical strategy that could most improve his/her appearance. We describe a method to predict the scoliotic trunk shape after a spine surgical intervention. The capability of our method was evaluated using real data of scoliotic patients. Results of the qualitative evaluation were very promising and a quantitative evaluation based on the comparison of the simulated and the actual postoperative trunk surface showed an adequate accuracy for clinical assessment. The required short simulation time also makes our approach an eligible candidate for a clinical environment demanding interactive simulations.

Analysis of scoliosis trunk deformities using ICA

This paper describes a method for analyzing scoliosis trunk deformities using Independent Component Analysis (ICA). Our hypothesis is that ICA can capture the scoliosis deformities visible on the trunk. Unlike Principal Component Analysis (PCA), ICA gives local shape variation and assumes that the data distribution is not normal. 3D torso images of 56 subjects including 28 patients with adolescent idiopathic scoliosis and 28 healthy subjects are analyzed using ICA. First, we remark that the independent components capture the local scoliosis deformities as the shoulder variation, the scapula asymmetry and the waist deformation. Second, we note that the different scoliosis curve types are characterized by different combinations of specific independent components.

Robust prediction of three-dimensional spinal curve from back surface for non-invasive follow-up of scoliosis

Spinal curvature progression in scoliosis patients is monitored from X-rays, and this serial exposure to harmful radiation increases the incidence of developing cancer. With the aim of reducing the invasiveness of follow-up, this study seeks to relate the three-dimensional external surface to the internal geometry, having assumed that that the physiological links between these are sufficiently regular across patients. A database was used of 194 quasi-simultaneous acquisitions of two X-rays and a 3D laser scan of the entire trunk. Data was processed to sets of datapoints representing the trunk surface and spinal curve. Functional data analyses were performed using generalized Fourier series using a Haar basis and functional minimum noise fractions. The resulting coefficients became inputs and outputs, respectively, to an array of support vector regression (SVR) machines. SVR parameters were set based on theoretical results, and cross-validation increased confidence in the systems performance. Predicted lateral and frontal views of the spinal curve from the back surface demonstrated average L2-errors of 6.13 and 4.38 millimetres, respectively, across the test set; these compared favourably with measurement error in data. This constitutes a first robust prediction of the 3D spinal curve from external data using learning techniques.