Chien-Chun Yang, Presenter: Nothing to Disclose

Mahesh Nagarajan, Abstract Co-Author: Nothing to Disclose

Markus B. Huber PhD, Abstract Co-Author: Nothing to Disclose

Jan Stefan Bauer MD, Abstract Co-Author: Nothing to Disclose

Felix Eckstein MD, Abstract Co-Author: Co-owner, Chondrometrics GmbH Co-founder, Chondrometrics GmbH CEO, Chondrometrics GmbH Consultant, Novartis AG Consultant, Merck KGaA Consultant, sanofi-aventis Group

Axel Wismueller MD, PhD, Abstract Co-Author: Nothing to Disclose

Thomas Baum MD, Abstract Co-Author: Nothing to Disclose

Eva-Maria Lochmueller MD, Abstract Co-Author: Nothing to Disclose

Julio Carballido-Gamio PhD, Abstract Co-Author: Nothing to Disclose

Thomas M. Link MD, PhD, Abstract Co-Author: Research funded, General Electric Company Research funded, InSightec Ltd

Biomechanical bone strength prediction in the proximal femur is a key component of osteoporosis diagnosis and associated fracture risk estimation. Our study proposes to use advanced integral geometry texture features derived from Minkowski Functionals for purposes of characterizing trabecular bone structure on multi-detector computed tomography (MDCT) images of femur specimens. Such novel topological feature vectors are subsequently compared with conventional measures of bone mineral density (BMD) in their ability to predict bone strength, which is achieved through support vector regression (SVR).

Axial MDCT images were acquired from 146 proximal femur specimens using a 16-row scanner and a calibration phantom. Spherical volumes of interest (VOI) were annotated in the femoral head (Huber et al., Radiology 2008) for BMD conversion and image analysis. VOIs of these BMD images were characterized through statistical moments as well as topological texture features derived from Minkowski Functionals. The specimens were then biomechanically tested by simulating a lateral fall on the greater trochanter, and failure load was recorded. All features were analyzed with multi-regression and SVR for predicting bone strength. The performance of different feature sets was compared using root-mean-square error (RMSE) and coefficient of determination (R2). A Wilcoxon signed-rank test was used to compare two RMSE distributions and test for statistically significant differences in performance.

The best prediction performance was observed with Minkowski Functional surface (RMSE = 0.939 ± 0.345, R2 = 0.544 ± 0.265) when analyzed with SVR, which was significantly lower than using mean BMD in conjunction with standard multi-regression analysis (RMSE = 1.075 ± 0.279, R2 = 0.417 ± 0.228) (p < 0.005).

Our results suggest that biomechanical strength prediction in the proximal femur can be significantly improved through topological characterization of trabecular bone micro-architecture, when used in conjunction with advanced machine learning techniques, such as support vector regression.

Complementing BMD characterization on MDCT images with advanced topological features and machine learning can contribute to improved diagnosis and disease progression monitoring in patients with osteoporosis.

Yang, C, Nagarajan, M, Huber, M, Bauer, J, Eckstein, F, Wismueller, A, Baum, T, Lochmueller, E, Carballido-Gamio, J, Link, T, Improving Bone Strength Prediction in Proximal Femur Specimens through Quantitative Characterization of Trabecular Micro-architecture with Minkowski Functionals and Support Vector Regression.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14010946.html