Demystifying machine learning approaches in digital bone imaging using microCT and HRpQCT

Micro-computed tomography (microCT) and high-resolution peripheral quantitative computed tomography (HRpQCT) generate three-dimensional digital images capturing bone structure and quality. Radiomic analytical approaches applied to these images extract quantitative measures of bone microarchitecture (e.g., bone volume and density). Automating and interpreting radiomics data using conventional image analysis techniques (e.g., bone segmentation) and statistical approaches is often inadequate due to their limited capacity to accommodate complex, nonlinear relationships. These limitations are especially apparent in bone research when integrating imaging outcomes with results from complementary analytical methods (e.g., biomechanics and histology) and experimental factors (e.g., clinical data). Machine learning (ML) offers opportunities in bone research by leveraging powerful computational tools; for example, to enhance bone spatial resolution, accelerate digital image segmentation, and reveal hidden patterns and relationships within high-dimensional bone data. Insights into key ML model inputs, which may be interpreted as primary biological phenotypes or therapeutic targets, can be revealed using standard (i.e., parametric and non-parametric analyses) or advanced statistical methods (i.e., dimensionality reduction and data integration). Overall, this narrative review has three main objectives: (1) to introduce current applications of ML in preclinical and clinical bone research using microCT and HRpQCT; (2) synthesize the interconnectedness of the field of bone and machine learning through user-friendly scientometric and bibliometric analyses and visualization using our novel software called SciNetX; and (3) to provide an accessible, high-level understanding of how ML models are developed and interpreted. These elements aim to provide a foundational guide to incorporating ML into bone research using digital imaging techniques.