Composite CT scoring system for Diffuse Idiopathic Skeletal Hyperostosis (DISH): a validated and reproducible...

Variability among observers remains a major barrier to standardizing imaging findings for spinal disease, hindering clinical diagnosis, research reproducibility, and the development of automated tools for assessment of spinal disease. Diffuse Idiopathic Skeletal Hyperostosis (DISH) is underdiagnosed in part due to the lack of consistent, CT-adapted scoring frameworks. We develop and validate a reproducible CT-based scoring system that integrates level-wise ossification scoring (Kuperus) with a structured severity classification (Holton) to capture DISH presence and bridging severity based on the number and distribution of bridged levels (scored 2 or 3). Five raters of varying experience levels (undergraduate students to an attending radiologist) scored 24 training scans; junior raters resolved discrepancies collaboratively, while senior raters reviewed independently. Pairwise comparisons showed perfect specificity (1.00) and moderate sensitivity (0.62-0.75) for DISH presence among less experienced raters compared to the attending radiologist. McNemar's tests showed no significant differences in pairwise disagreement rates for DISH presence, while paired Wilcoxon tests revealed significant differences in bridged level counts, with less experienced raters underestimating the number of bridged levels. Scoring time was 2-3 minutes per scan following training. Subsequently, two independent raters scored 614 thoracolumbar CT scans. Interobserver reliability was excellent for DISH presence (Gwet's AC2 = 0.957 [95% CI: 0.936-0.977]), bridged level count (ICC = 0.933 [95% CI: 0.922-0.943]), and bridging severity (AC2 = 0.950 [95% CI: 0.931-0.969]). Level-wise scores also showed high agreement (median AC2 = 0.881, range: 0.862-0.937), with discrepancies occurring primarily due to transitional anatomy and mildly ossified levels. This CT-based scoring framework enables reliable and reproducible labeling of DISH presence and severity across raters of varying experience levels. Its structured format supports scalable image annotation and minimizes label noise, providing a robust foundation for training computer vision models to automate detection and stratification of DISH in spinal imaging.