Automated segmentation of the sacro-iliac joints, posterior spinal joints and discovertebral units on low-dose computed tomography for Na[18F]F PET lesion detection in spondyloarthritis patients

Abstract

Purpose: Spondyloarthritis (SpA) is a chronic inflammatory rheumatic disease which involves the axial skeleton. Quantitative sodium fluoride-18 (Na[¹⁸F]F) PET/CT is a new imaging approach promising for accurate diagnosis and treatment monitoring by assessment of molecular bone pathology in SpA. Detection of Na[¹⁸F]F PET positive lesions is time-consuming and subjective, and can be replaced by automatic methods. This study aims to develop and validate an algorithm for automated segmentation of the posterior spinal joints, sacro-iliac joints (SIJs) and discovertebral units (DVUs) on low-dose computed tomography (LDCT), and to employ these segmentations for threshold-based lesion detection.

Methods: Two segmentation methods were developed using Na[¹⁸F]F PET/LDCT images from SpA patients. The first method employed morphological operations to delineate the joints and DVUs, while the second used a multi-atlas-based approach. The performance and reproducibility of these methods were assessed on ten manually segmented LDCTs using average Hausdorff distance (HD) and dice similarity coefficient (DSC) for DVUs and SIJs, and mean error distance for the posterior joints. Various quantitative PET metrics and background corrections were compared to determine optimal lesion detection performance relative to visual assessment.

Results: The morphological method achieved significantly better DSC (0.82 (0.73-0.88) vs. 0.74 (0.68-0.79); p < 0.001) for all DVUs combined compared to the atlas-based method. The atlas-based method outperformed the morphological method for the posterior joints with a median error distance of 4.00 mm (4.00-5.66) vs. 5.66 mm (4.00-8.00) (p < 0.001). For lesion detection, the atlas-based segmentations were more successful than the morphological method, with the most accurate metric being the maximum standardized uptake value (SUVmax) of the lesional Na[¹⁸F]F uptake, corrected for the median SUV (SUVmedian) of the spine, with an area under the curve of 0.90.

Conclusion: We present the first methods for detailed automatic segmentation of the posterior spinal joints, DVUs and SIJs on LDCT. The atlas-based method is the most appropriate, reaching high segmentation performance and lesion detection accuracy. More research on the PET-based lesion segmentation is required, to develop a pipeline for fully automated lesional Na[¹⁸F]F uptake quantification.

Keywords: Artificial intelligence-based segmentation; Automated lesion detection; PET quantification; Rheumatic disease; Spinal bone formation.

Fig. 1

Morphological segmentation of the vertebral bodies (VBs). a, b Axial slice with initial segmentation of C4 (Green) (a) and L1 (Blue) (b) with the spinal cord (Purple). c Most anterior point of the spinal cord (Red). d Central point of spinal cord (Red). e, f VB segmentation, only containing pixels anterior of the spinal cord point

Fig. 2

Atlas-based method. a Atlas creation. b Atlas-based segmentation. VB: Vertebral body, IVD: Intervertebral disk

Fig. 3

Composition of the disco vertebral unit segmentation. VB: Vertebral Body, DVU: Discovertebral unit

Fig. 4

Segmentation results for the intervertebral disks (a–d) and the sacro-iliac joints (e–h). a, e Initial segmentation. b, f Manual segmentation. c, f Atlas-based segmentation. d, h Morphological segmentation

Fig. 5

Boxplots of the segmentation performance metrics for the atlas-based method (Red) (A) and the morphological method (Blue) (M), compared to the manual segmentation. a Dice similarity coefficients (DSC). b Average Hausdorff distance (HD). c Error distance between joint centers. DVU: Discovertebral Unit, SIJ: Sacro-iliac joint, CVJ: Costovertebral joint, CTJ: Costotransverse joint

Fig. 6

Receiver operator curves of the different metrics applied for optimal lesion detection performance. a Comparison of SUVmax, SUVmean, and SUVpeak for both the atlas-based method (A) and the morphological method (M). b Comparison of the best-performing metric SUVmax in the atlas-based segmentation corrected by the SUVmedian of different possible background structures