Introduction: To determine curve type, Lenke classification for AIS uses strict cut-off values on radiological measurements such as Cobb angles which are known to have significant inter-observer variability. There is a documented variability in surgical treatment of AIS, yet the influence of curve types on that variability has not yet been studied.
Objectives: To use an automated method to classify AIS patients using radiological measurements. Our working hypothesis is that Kohonen Self-Organizing Maps (SOM) can avoid limitations seen with classification using strict criteria. It can also highlight treatment variability depending on curve types.
Methods: Pre-operative Cobb angles from 1801 surgically treated AIS cases were inputted into a neural network to generate a SOM onto which Lenke classes and fusion levels were transposed. Geometric validation of the map using threedimensional reconstruction was done and Kappa statistics were used to evaluate treatment variability.
Results: SOM classify scoliotic spines with a distribution gradient for each of the parameters inputted. The levels of fusions were only homogeneous in single thoraco-lumbar curves with a kappa value of 1.0 . 71 three-dimensional reconstruction of scoliotic spines were mapped on the kohonen map showing conservation of geometrical neighbouring.
Conclusion: SOM can efficiently classify AIS while respecting neighbouring of similar scoliotic spines. There is ubiquitous variability in surgical treatment of AIS with the exception of single thoraco-lumbar/lumbar curves.
Significance: Such classifications will allow us to better query large database to lookup for similar cases while eliminating the limitations imposed by classifications using rigid criteria.