Using a fully automated deep reasoning and learning model to distinguish bone health on computed tomography for adolescent idiopathic scoliosis (AIS) and non-AIS patients: a pilot study.
Authors
Affiliations (7)
Affiliations (7)
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN, USA.
- Orthopedic Surgery Artificial Intelligence Laboratory, Mayo Clinic, Rochester, MN, USA.
- 3D Anatomic Modelling Laboratory, Department of Radiology, Mayo Clinic, Rochester, MN, USA.
- Limb and Musculoskeletal Regenerative Research Laboratory, Mayo Clinic, Rochester, MN, USA.
- Department of Orthopedic Surgery, Mayo Clinic, Rochester, MN, USA. [email protected].
- Orthopedic Surgery Artificial Intelligence Laboratory, Mayo Clinic, Rochester, MN, USA. [email protected].
- Mayo Clinic Children's, Mayo Clinic, 200 1 st Street SW, Rochester, MN, 55905, USA. [email protected].
Abstract
Accurate evaluation of spinal bone density is essential for assessing bone health, particularly in preoperative planning for spine surgery patients. Conventional manual methods for measuring bone density are time-consuming and rely on a single slice, limiting precision and consistency. We hypothesize that a deep reasoning and learning-based (DR-AI) model can fully automate volumetric spinal bone density assessment, differentiating clinically favorable versus poor bone quality. A cross-sectional study was conducted using preoperative low-dose CT scans from 20 patients: 10 with adolescent idiopathic scoliosis (AIS) with favorable intraoperative bone quality assessment (Grade A) and 10 with neuromuscular (NM) scoliosis or neurofibromatosis (NF) demonstrating poor intraoperative bone quality (Grade D). The DR-AI model segmented and quantified volumetric Hounsfield Units (vHU) for each vertebra from T1 to L5. Multivariate Analysis of Variance (MANOVA), adjusted for age and gender, was used to compare vHU values between the cohorts. Polyserial correlations assessed association between surgeon-assigned grades and vHU measurements. MANOVA demonstrated differences in vHU between AIS and non-AIS cohorts from T3 to L5 (all p < 0.05). On average, AIS patients had higher mean vHU values (251.1 ± 37.9) than non-AIS patients (190.6 ± 54.9). vHU values demonstrated moderate-to-very strong correlations with surgeons' intraoperative assessment, ranging from ρ = 0.5 at T2 to ρ = 0.9 at L5. The DR-AI model may represent a promising tool for preoperative planning, offering a faster, more objective approach for automated volumetric evaluation of vertebral bone characteristics through vHU. Future research should validate this technology in larger, more diverse cohorts. Level IV-Cross sectional study.