ESPE Abstracts (2015) 84 FC2.6

A Contextual Feature-Based Recognition Approach to Quantify Trabecular Microstructure Using 1.5T Axial-MRI: An Innovative Methodology

Paul Dimitria, Karim Lekadirc, Corné Hoogendoornc, Paul Armitageb, Elspeth Whitbyb, David Kinga, Richard Eastellb & Alejandro Frangib


aSheffield Children’s Hospital, Sheffield, South Yorkshire, UK; bThe University of Sheffield, Sheffield, South Yorkshire, UK; cUniversitat Pompeu Fabra, Barcelona, Spain


Background: In-vivo skeletal MRI imaging remains challenging due to the extremely short MR relaxation times (<1 ms) of protons bound to water in bone. However, each MRI sequence contains properties identifiable through feature-based recognition, highlighting characteristics relating to skeletal configuration. We thus present a novel statistical method for clinical 1.5 Tesla (T) MRI in quantifying trabecular microstructure and use HRpQCT to determine its accuracy.

Objective: To assess fifteen contextual image-based features of trabecular bone relating to pattern fragmentation, repeatability, complexity, and statistical variability of multiple axial-plane MRI sequences.

Method: We compared HRpQCT and 1.5T MRI scans of the non-dominant ultradistal tibia in 96 13–16 year olds. Participants underwent two of the following axial-MRI sequences: T1-weighted fast spin echo, T2-weighted fast spin echo, T2*-weighted gradient echo, FIESTA, ultrashort time echo (UTE), UTE High Resolution. By relating HRpQCT-derived trabecular parameters to contextual image features contained within axial-MRI sequences we developed a statistical prediction model designed to predict trabecular microstructural parameters. Image descriptors included statistical variability (mean intensity, standard deviation, skewness, and kurtosis), pattern repeatability (using grey level co-occurrence matrices), and pattern complexity (using run-length analysis and fractal dimension). Kernel partial least squares defined an optimal non-linear predictor model from the data relating MRI sequences to HRpQCT parameters. By using the MRI sequences as the input of the prediction model: ((ymripredicted−yhrpqct/yhrpqct)×100), percentage prediction errors for trabecular thickness, spacing and number were calculated.

Results: FIESTA and UTE-HR image sequences demonstrated the highest accuracy in predicting all three trabecular parameters (12.0±3.4%, 12.1±4.5% respectively). T2w and T2×w most accurately predicted trabecular thickness (mean prediction error- 9.5%) and trabecular number (7.5%), respectively. T1w most accurately predicted trabecular spacing (7.4%).

Conclusion: 1.5T MRI sequences can predict trabecular number, spacing, and thickness to within 10% of the values derived from HRpQCT using the established model, demonstrating the future potential of clinical 1.5T MRI in assessing trabecular bone.

Funding: The British Society of Paediatric Endocrinology and Diabetes. The Sheffield Children’s Hospital Charity, UK.

Article tools

My recent searches

No recent searches.