Nondisease Sources of Radiomic Texture Variation on Plain Radiographs: A Cross-Sectional Analysis of Anatomical Region, Projection, and Resolution in a Public Fracture Archive.
Authors
Affiliations (8)
Affiliations (8)
- Department of Surgery, Kansas City University School of Medicine, Kansas City, MO, USA. [email protected].
- Freeman Health System, Joplin, MO, USA. [email protected].
- Atrium Health Wake Forest Baptist Medical Center, Winston-Salem, NC, USA. [email protected].
- Department of Surgery, Kansas City University School of Medicine, Kansas City, MO, USA.
- Freeman Health System, Joplin, MO, USA.
- Atrium Health Wake Forest Baptist Medical Center, Winston-Salem, NC, USA.
- Department of Physical Medicine and Rehabilitation, University of Missouri School of Medicine, Columbia, MO, USA.
- Wake Forest University School of Medicine, Winston-Salem, NC, USA.
Abstract
Radiomic texture features from plain radiographs feed machine learning models on the assumption that they reflect tissue rather than anatomy or acquisition. In a public multihospital fracture archive, we quantified how anatomical region and radiographic projection, each adjusted for the other and for image resolution, shift 17 radiomic features, how much feature variance resolution explains, and a bony superimposition contrast defined within region. This retrospective cross-sectional secondary analysis used FracAtlas (4083 musculoskeletal radiographs from three hospitals); 4024 images were analyzed. For each z-scored feature (10 first-order, 7 Gy-level co-occurrence matrix [GLCM]), linear models estimated adjusted differences (β, SD units), and resolution by variance explained (R<sup>2</sup>), with Benjamini-Hochberg false discovery rate correction within each factor. Superimposition (hip and shoulder vs hand and leg; n = 3630), a grouping of regions, was analyzed as a pooled regional contrast rather than an independent factor. Region shifted 16 of 17 features (radiomic uniformity, hand vs leg - 0.66 SD [95% CI - 0.72 to - 0.59]). Resolution explained up to 16.8% of single-feature variance and changed the fracture log-odds coefficient by more than 15% for 7 of 10 interpretable features. Projection had the smallest effects (radiomic 10th percentile, lateral vs frontal - 0.32 SD). The superimposition contrast produced the largest shifts (GLCM energy - 2.26 SD [95% CI - 2.36 to - 2.16]). Radiomic texture on plain radiographs varies substantially with anatomy and acquisition, independently of fracture status. Radiograph radiomics studies should record and adjust for region, projection, and resolution, report region-specific estimates, and validate across sites.