A deep learning-based framework for standardized analysis of trabecular bone compartments from micro-CT imaging data in the mouse tibia

Comprehensive Summary

This study explores a new deep learning alternative to high-resolution micro-computed tomography (micro-CT) imaging analysis on trabecular bone. To conduct the study, a deep learning framework was trained and tested on 40 annotated bone scans of the epiphyseal-metaphyseal region in a mouse tibia. The model achieved mean F1-scores of 0.96 for epiphyseal bone, 0.95 for the growth plate, 0.92 for the primary spongiosa, and 0.99 for the secondary spongiosa across all datasets. On an external dataset, the model achieved mean F1-scores of 0.99, 0.97, 0.92, and 1.00 for these same regions, respectively.

Outcomes and Implications

Current manual segmentation methods of micro-CT trabecular bone analysis are susceptible to inconsistencies and non-standardized definitions of volumes of interest (VOIs). This affects reproducibility and can lead to misleading statistical interpretations. With consistent, automated analysis of trabecular bone, preclinical skeletal research would improve significantly, bone disease progression could be monitored more intensely, and the efficacy of therapeutic interventions could be evaluated. Although the author does not explicitly comment on the timeline for clinical implementation, studies on human bone would be required before generalizing this workflow to patients.