Local fragility index and Radiomics-lite modeling of CT-Derived Hounsfield units for osteoporosis risk assessment in
Akihiko Hiyama1, Daisuke Sakai2, Hiroyuki Katoh2
1Department of Orthopaedic Surgery, Tokai University School of Medicine, Isehara, Kanagawa, Japan. a.hiyama@tokai.ac.jp.
Study Design:
Retrospective cohort study.
Purpose:
To develop and validate the Local Fragility Index (LFI), a novel CT-derived measure that integrates the lowest vertebral Hounsfield Unit (HU) (Min HU), inter-level variability (SD), and cranio-caudal slope to quantify vertebral heterogeneity and improve osteoporosis risk assessment.
Methods:
We retrospectively analyzed 172 patients who underwent spinal surgery. Severe adult spinal deformity (Cobb angle > 30°) was excluded to ensure measurement consistency. HU values from L1-L4 were obtained on preoperative CT. Min HU was inverted (- Min HU) to emphasize fragility, SD was calculated across levels, and the cranio-caudal slope of HU values was estimated by linear regression across L1-L4. Each parameter was standardized using Z-scores, and LFI was defined as z(- Min HU) + z(SD) + z(- Slope). The composite score was then further standardized to generate a normalized index (LFI_z), which was used for all statistical analyses. Osteoporosis was defined as the lowest DXA T-score ≤ - 2.5. Diagnostic performance was assessed by ROC analysis, logistic regression, decision curve analysis (DCA), and category-free net reclassification improvement (cfNRI).
Results:
Min HU was significantly lower in osteoporotic patients (p < 0.001) and correlated with the lowest DXA T-score (r = 0.52). LFI_z showed comparable discrimination to Mean HU (AUC 0.664 vs. 0.690), while their combination achieved higher accuracy (AUC 0.701). In adjusted models, LFI_z remained an independent predictor of osteoporosis (OR 1.9 per SD, p < 0.01), with improved reclassification (cfNRI + 0.19, 95% CI 0.05-0.32). The Radiomics-lite + Clinical model (HU quantiles, variability metrics, and demographics) provided the best performance (AUC 0.824) and the greatest net clinical benefit across 10-40% thresholds.
Conclusions:
The LFI captures vertebral heterogeneity and complements mean HU, thereby enhancing osteoporosis risk stratification. Building on this framework, a simplified Radiomics-lite + Clinical model further improved discriminative performance and decision-analytic utility, offering a scalable approach for CT-based assessment in patients undergoing spine surgery. This approach may directly inform surgical planning (e.g., screw augmentation, extended fusion levels) and perioperative management of osteoporosis. As this outcome-agnostic, hypothesis-generating analysis was benchmarked to DXA rather than clinical endpoints, prospective validation against fractures and instrumentation-related complications is needed before clinical use.
