Automatic segmentation of the spinal joints and intervertebral disks on low-dose computed tomography: an atlas-based approach

Abstract

Introduction: Spondyloarthritis (SpA) belongs to the chronic inflammatory rheumatic diseases, and primarily affects the axial skeleton. Quantitative 18F-NaF PET/CT is a new imaging approach that shows promise for accurate diagnosis and treatment assessment. Manual segmentation of low-dose computed tomography (LDCT) for quantitative feature extraction is time-consuming and subjective, and can be replaced by automatic methods. This study aims to develop and validate an automatic algorithm for the segmentation of the spinal joints and intervertebral disks (IVD’s) on LDCT using two different approaches.
Methods: Two methods for spinal structure segmentation were developed and compared. Both methods used segmentations of bony structures obtained from the TotalSegmentator algorithm. The first method employed morphological dilation and erosion operations to localise the joints and IVD’s, while the second method used a multi-atlas-based method approach with partial atlases and corresponding manually segmented labelmaps. The performance of the methods was assessed on ten manually segmented LDCT’s using sensitivity, and maximum and average Hausdorff distance (HD) for IVD’s and the sacroiliac joints (SIJ) and mean error distance for the smaller joints. The reproducibility of the methods was evaluated using a set of 20 LDCT test-retest images.
Results: The atlas-based method achieved significantly better maximum HD (8.45 (1.80) vs. 9.64 (5.83) (p = 0.002)) and sensitivity (0.79 (0.22) vs. 0.61 (0.30) (p < 0.001)) for all IVD’s combined compared to the morphological method. The atlas-based method also outperformed the morphological method for the facet joints, costovertebral joints and costotransverse joints with a mean error distance of 4.71 mm (2.72) vs 6.90 mm (4.80) (p < 0.001). For the thoracic IVD’s the morphological method showed significantly better average HD (1.48 (1.03) vs. 1.72 (0.53) (p = 0.018)) and maximum HD (6.97 (3.36) vs. 8.22 (1.66) (p < 0.001)) than the atlas-based method. In the reproducibility assessment on the test-retest scans, the atlas-based method outperformed the morphological method for all metrics and structures, with average HD’s well below the voxel resolution (< 2 mm).
Conclusion: We present the first methods for automatic segmentation of the spinal structures on LDCT. The atlas-based method seems to be the most suitable algorithm, achieving average HD’s below the voxel size, and maximum errors below one centimetre. However, it is dependent on accurate segmentation by the TotalSegmentator algorithm. Further research is warranted to investigate the influence of the segmentation results on the extraction of quantitative PET information.