Spatial-resolution Robustness of CT-Based Vertebral Fragility Screening in Potentially War-affected Individuals

Abstract

The paper is devoted to preserving the accuracy of machine-learning models for computed tomography based vertebral fragility screening under degraded image acquisition, typical of mobile and field-deployed scanners in war-affected regions. The paper proposes a paired evaluation protocol, in which a model trained on clean features is applied twice to the same held-out vertebrae, which is based on controlled synthetic degradation of the same computed tomography volumes: additive noise (15, 30, 50 HU), in-plane Gaussian blur (σ = 0.5, 1, 2 mm) and slice-thickness coarsening (2, 3, 5, 7 mm). The paper develops a comparative assessment of four imaging representations of vertebral bone on two independent cohorts (106 patients, 478 vertebrae and 126 patients, 589 vertebrae), the paired difference of the area under the ROC curve being estimated by a patient-clustered bootstrap. Additive noise produced no dose-consistent effect. In-plane blur reduced the area under the curve of the radiomics representation by 0.096 (95 % CI –0.165, –0.025) and 0.112 (–0.169, –0.053) in the two cohorts, whereas the mean-attenuation representation stayed within the null interval at every condition (0.732 and 0.665 on clean data). The results of the work can be used to select acquisition-robust imaging biomarkers for opportunistic screening of war-affected and displaced individuals, whose mean age matches that of the cohorts studied and who are scanned wherever equipment is available.