Adapting super-resolution reconstruction for skeletal analysis of clinical computed tomography data

Abstract Clinical computed tomography (CT) datasets are increasingly common in skeletal research, yet archived retrospective datasets with thick-slice, anisotropic reconstructions are more commonly available to researchers than the original high-resolution scans. Models reconstructed from these suboptimal scans can produce distorted measurements and irregular surfaces. This study evaluates a super-resolution reconstruction (SRR) framework for generating skeletal models with improved surface smoothness and continuity from multiple orthogonal, thick-slice CT stacks. Archived CT scans of long bones from 30 individuals (0–16 years) were collected from National Taiwan University Hospital. For each individual, 3D models were generated from both the original thick-slice stacks and SRR-processed volumes. Linear measurements were compared with those taken directly from high-resolution picture archiving and communication system (PACS) renderings to assess metric consistency. Geometric similarity between original thick-slice models and SRR-reconstructed models was quantified by signed surface deviation and Dice similarity coefficient. Because both model types were derived from the same image stacks, this geometric comparison further quantifies consistency between reconstruction approaches. SRR-reconstructed models showed lower measurement error and greater agreement with the PACS-derived measurements than the thick-slice models. Surface geometry was generally consistent across model types, with localized deviations concentrated at metaphyseal regions. SRR processing also produced smoother surfaces with a clearer separation of fusing elements. These results demonstrate that SRR is a practical tool to improve surface continuity in virtual skeletal models derived from suboptimal clinical imaging.