Neural-network Detection of Thoracic Curve Severity in Scoliosis

Torso deformity is often the primary complaint of a patient with scoliosis. Non-invasive detection of progression of the underlying scoliosis from torso deformity would reduce the cancer risk associated with spinal X-rays (Levy et al., 1996) and may define the 3-D scoliosis deformity more completely and with greater sensitivity to patient concerns than simply measuring the Cobb angle on a plain X-ray. While studies to date have shown only moderate torso-spine correlations (Stokes & Moreland, 1989), in a recent study using an artificial neural network (ANN) to analyze 3D fulltorso surface topographic data in 65 scoliotic patients we were able to predict the maximal spinal Cobb angle within 6.5° in 63% of patients (Jaremko et al., 2001). A clinical test should be both sensitive (few false negative results) and specific (few false positive results). Here, we measured the ability of the torso-scanner/ANN system to distinguish between thoracic spinal curves of varying severity by determining the sensitivity and specificity of ANN predictions of the thoracic Cobb angle.