Voxel-Level Forecasting of Apparent Vertebral Bone Loss from Routine CT Using a 3D Convolutional Neural Network

Existing computed tomography (CT) fracture-risk models typically output one scalar score per vertebra without localising the predicted change. We built and internally validated a three-dimensional (3D) convolutional network forecasting, from one baseline acquisition, a per-voxel map of apparent inter-scan bone-mask loss per lumbar vertebra. Paired CT studies from 65 patients yielded 298 baseline-unfractured vertebra patches (L1 to L5); every patient sustained at least one reader-adjudicated incident fracture, so the cohort is outcome-enriched. The endpoint was apparent vertebral-body mask loss between scans after rigid alignment of automated masks. A multi-task 3D residual U-Net with vertebral-level conditioning was compared with a gradient-boosted classifier on the same features and folds, under five-fold patient-grouped cross-validation with clustered bootstrap intervals. Pooled precision-recall area was 0.445 vs. 0.439 for the reference (+0.005, 95% confidence interval −0.018 to +0.028) and pooled area under the curve (AUC) was 0.815. At matched predicted volume, the network map was less fragmented (2.3 vs. 5.3 components per 1000 voxels), and the endpoint separated incident fractures (AUC 0.865). Voxel-level forecasting of apparent inter-scan vertebral change is feasible on routine CT; the contribution is the paired-scan forecasting framework and the spatial-profile difference, not better discrimination. The endpoint is an algorithmic surrogate requiring voxel-level expert adjudication and external validation before clinical use.