Application of machine learning classifiers for microcomputed tomography data assessment of mouse bone microarchitecture
Jennifer C. Coulombe,
Zachary K. Mullen,
Maureen E. Lynch,
Louis S. Stodieck,
Virginia L. Ferguson
Affiliations
Jennifer C. Coulombe
Department of Mechanical Engineering, UCB 427, University of Colorado, Boulder, CO 80309, United States of America; BioFrontiers Institute, UCB 596, University of Colorado, Boulder, CO 80309, United States of America
Zachary K. Mullen
Laboratory for Interdisciplinary Statistical Analysis / Department of Computer Science, UCB 427, University of Colorado, Boulder, CO 80309, United States of America
Maureen E. Lynch
Department of Mechanical Engineering, UCB 427, University of Colorado, Boulder, CO 80309, United States of America; BioFrontiers Institute, UCB 596, University of Colorado, Boulder, CO 80309, United States of America
Louis S. Stodieck
Aerospace Engineering Sciences / BioServe Space Technologies, UCB 429, University of Colorado, Boulder, CO 80309, United States of America
Virginia L. Ferguson
Department of Mechanical Engineering, UCB 427, University of Colorado, Boulder, CO 80309, United States of America; BioFrontiers Institute, UCB 596, University of Colorado, Boulder, CO 80309, United States of America; Aerospace Engineering Sciences / BioServe Space Technologies, UCB 429, University of Colorado, Boulder, CO 80309, United States of America; Corresponding author at: Department of Mechanical Engineering, UCB 427, University of Colorado, Boulder, CO 80309, USA.
The current standard approach for analyzing cortical bone structure and trabecular bone microarchitecture from micro-computed tomography (microCT) is through classic parametric (e.g., ANOVA, Student's T-test) and nonparametric (e.g., Mann-Whitney U test) statistical tests and the reporting of p-values to indicate significance. However, on their own, these univariate assessments of significance fall prey to a number of weaknesses, including an increased chance of Type 1 error from multiple comparisons. Machine learning classification methods (e.g., unsupervised, k-means cluster analysis and supervised Support Vector Machine classification, SVM) simultaneously utilize an entire dataset comprised of many cortical structure or trabecular microarchitecture measures, thus minimizing bias and Type 1 error that are generated through multiple testing. Through simultaneous evaluation of an entire dataset, k-means and SVM thus provide a complementary approach to classic statistical analysis and enable a more robust assessment of microCT measures.
