Comparison of Segmentation Methods in Analysis of MR and CT Images of Pediatric Spine

Scoliosis is the most common spinal deformity in children. Only early treatment during spinal growth can significantly reduce the associated problems caused by the deformity in adults. The aim of this study is to use a spine model to numerically simulate the changes in spinal stresses during correction of congenital deformity by vertebral osteotomy. In the first stage, CT imaging was used as a reference to obtain correctly segmented vertebral groups due to the low quality of MRI image data. Registration techniques were optimized to process all MRI and CT image sequences. An SVM classifier was used with Dice coefficients of 0.98 for CT and 0.95, 0.97, 0.91 and 0.92 for T1 hard, T2 hard, T1 soft and T2 soft, respectively. In the next phase of the project, deep learning algorithms were used to obtain MRI segmentation. Two different segmentation algorithms were proposed using the U-Net network. Standard and patchwise approach with rotational averaging for both CT and MRI dataset. The standard segmentation produced more accurate results with a Dice coefficient of 0.96 for the CT dataset and 0.94 for the MRI dataset. The patchwise method provided slightly better results when processing the actual dataset containing the new data acquired by our MRI scanner. With the smaller MRI dataset, we achieved comparable Dice coefficients in both datasets. The presented results suggest the possibility of using CT and even MR imaging exclusively for spine segmentation if visualization of surrounding tissues and automatic 3D spine modeling is desired.