A Laplace-Hamming Binarization Approach for Second-Generation HR-pQCT Rescues Fine Feature Segmentation

Abstract

Although second-generation high-resolution peripheral quantitative computed tomography (XCTII) provides the highest-resolution in vivo bone microstructure assessment, the manufacturer's standard image processing protocol omits fine features in both trabecular and cortical compartments. To optimize fine structure segmentation, we implemented a binarization approach based on a Laplace-Hamming (LH) segmentation and documented the reproducibility and accuracy of XCTII structure segmentation using both the standard Gaussian-based binarization and the proposed LH segmentation approach. To evaluate reproducibility, 20 volunteers (9 women, 11 men; aged 23-75 years) were recruited, and three repeat scans of the radii and tibias were acquired using the manufacturer's standard in vivo protocol. To evaluate accuracy, cadaveric structure phantoms (14 radii, 6 tibias) were scanned on XCTII using the same standard in vivo protocol and on μCT at 24.5 μm resolution. XCTII images were analyzed twice-first, with the manufacturer's standard patient evaluation protocol and, second, with the proposed LH segmentation approach. The LH approach rescued fine features evident in the grayscale images but omitted or overrepresented (thickened) by the standard approach. The LH approach significantly reduced error in trabecular volume fraction (BV/TV) and thickness (Tb.Th) compared with the standard approach; however, higher error was introduced for trabecular separation (Tb.Sp). The LH approach improved the correlation between XCTII and μCT for cortical porosity (Ct.Po) and significantly reduced error in cortical pore diameter (Ct.Po.Dm) compared with the standard approach. The LH approach resulted in improved precision compared with the standard approach for BV/TV, Tb.Th, Ct.Po, and Ct.Po.Dm at the radius and for Ct.Po at the tibia. Our results suggest that the proposed LH approach produces substantially improved binary masks, reduces proportional bias, and provides greater accuracy and reproducibility in important outcome metrics, all due to more accurate segmentation of the fine features in both trabecular and cortical compartments. © 2023 The Authors. Journal of Bone and Mineral Research published by Wiley Periodicals LLC on behalf of American Society for Bone and Mineral Research (ASBMR).

Keywords: ACCURACY; GAUSSIAN BINARIZATION; HR-PQCT; LAPLACE-HAMMING BINARIZATION; REPRODUCIBILITY.

Figure 1.

The custom structure phantom: each phantom consists of a 7-cm-diameter cylinder of soft-tissue equivalent polymer resin; each cylinder contains five 1-cm-thick cadaveric bone embedments, sectioned at a location corresponding to the standard location for an in vivo HR-pQCT exam. A representative image for one bone section (radius) is shown on the right [figure taken with permission from (Burghardt et al., 2013)].

Figure 2.

Grayscale, standard XCTII and the proposed LH segmentation approach for an in vivo radius (first row) and a phantom radius section (second row) showing that the LH approach captures fine features with greater fidelity.

Figure 3.

Grayscale, standard XCTII and the proposed LH segmentation approach for an in vivo tibia (first row) and a phantom tibia section (second row) showing that the LH approach rescues cortical pores lost in the standard segmentation.

Figure 4.

Regression and Bland-Altman plots for BV/TV (A-C), Tb.Th (D-F) and Tb.N (G-I) assessed using standard XCTII and LH approach for accuracy comparisons. On regression plots, the dashed line indicates the line of unity. On Bland-Altman plots, the solid line indicates the mean difference and the dashed lines indicate the 95% limits of agreement.

Figure 5.

Regression and Bland-Altman plots for Ct.Po (A-C), and Ct.Po.Dm (D-F) assessed using standard XCTII and LH approach for accuracy comparisons. On regression plots, the dashed line indicates the line of unity. On Bland-Altman plots, the solid line indicates the mean difference and the dashed lines indicate the 95% limits of agreement.

Figure 6.

Regression and Bland-Altman plots for BV/TV (A-B), Tb.Th (C-D), Tb.N (E-F), Ct.Po (G-H), and Ct.Po.Dm (I-J) assessed using standard XCTI approach for accuracy comparisons. On regression plots, the dashed line indicates the line of unity. On Bland-Altman plots, the solid line indicates the mean difference and the dashed lines indicate the 95% limits of agreement.