Protocol for an Anatomy-Constrained Simulation Framework for DXA-Grounded Osteoporosis Screening

Abstract

Opportunistic osteoporosis screening from conventional radiographs could identify patients who maybenefit from confirmatory dual-energy X-ray absorptiometry (DXA), but development is constrained bylimited paired radiograph-DXA data and under-representation of clinically important bone-density strata.This study proposes an anatomy-constrained simulation framework for generating lumbar radiographs withcontrollable bone mineral density (BMD) and morphology, coupled to a multi-task screening model. Theframework represents lumbar vertebrae L1-L4 as an ordered anatomical graph, conditions image synthesison DXA-derived BMD and osteoporosis status, and constrains generated images using vertebral geometry,cortical boundaries, and inter-vertebral consistency. A downstream network jointly estimates continuousBMD, classifies normal bone density, osteopenia, and osteoporosis, and produces an uncertainty-calibratedrecommendation for confirmatory DXA. Evaluation is specified on the openly available LUMOS dataset,which contains 803 patients, 1,620 anteroposterior or lateral lumbar radiographs, DXA measurements, and280 CT examinations. Comparisons include real-only, synthetic-only, unconstrained augmentation, andanatomy-constrained augmentation, with patient-level cross-validation and calibration assessment. Theprincipal hypothesis is that anatomically constrained simulation improves minority-class sensitivity androbustness without degrading calibration. This work is positioned as a simulation methodology andscreening study rather than evidence for autonomous diagnosis. Clinical deployment would requireexternal, prospective, and demographically broader validation.