Fully automated segmentation of callus by micro-CT compared to biomechanics

Abstract

Background: A high percentage of closed femur fractures have slight comminution. Using micro-CT (μCT), multiple fragment segmentation is much more difficult than segmentation of unfractured or osteotomied bone. Manual or semi-automated segmentation has been performed to date. However, such segmentation is extremely laborious, time-consuming and error-prone. Our aim was to therefore apply a fully automated segmentation algorithm to determine μCT parameters and examine their association with biomechanics.

Methods: The femura of 64 rats taken after randomised inhibitory or neutral medication, in terms of the effect on fracture healing, and controls were closed fractured after a Kirschner wire was inserted. After 21 days, μCT and biomechanical parameters were determined by a fully automated method and correlated (Pearson's correlation).

Results: The fully automated segmentation algorithm automatically detected bone and simultaneously separated cortical bone from callus without requiring ROI selection for each single bony structure. We found an association of structural callus parameters obtained by μCT to the biomechanical properties. However, results were only explicable by additionally considering the callus location.

Conclusions: A large number of slightly comminuted fractures in combination with therapies that influence the callus qualitatively and/or quantitatively considerably affects the association between μCT and biomechanics. In the future, contrast-enhanced μCT imaging of the callus cartilage might provide more information to improve the non-destructive and non-invasive prediction of callus mechanical properties. As studies evaluating such important drugs increase, fully automated segmentation appears to be clinically important.

Keywords: 3D structural parameters; Biomechanics; Comminuted fracture; Fracture healing; Fully automated segmentation; Micro-CT (μCT); Multi-fragmented fracture; Trabecular bone.

Fig. 1

2D axial μCT grey-value image of the original cortical bone, callus and air/bone marrow (a, e), conventionally contoured images according to the method described by Namayn et al. (b) and Morgan et al. (c) and a corresponding fully automated segmented image (d, callus = red, original bone = white, marrow/air = black). A single contour adjacent to the callus distorted the evaluation of the volume-dependent callus parameters (e.g. BV/TV and BMD) because TV enclosed (in addition to the callus) all structures within the contour, such as the cortical bone and air/marrow. Therefore, the adjacent contouring is not necessary (b, f). Contouring of both the outer callus (green line) and outer cortical bone (red line) enables all callus parameters to be determined (between the lines) in an area not corresponding to the fracture gap (c). The latter was not possible within the fracture gap, especially if slight comminution was present (g). The enclosed air and unmineralised tissue would contribute to considerable underestimation of the volume-dependent parameters, and no measurement of the endosteal callus would be possible. Fully automated segmentation required no adjacent contouring of the ROI and enabled, even within the fracture gap, the separation of all islands of callus and cortical bone as well as determination of the non-volume-dependent parameters (h; for ROI, see Fig. 3). Upper line: the same images of an area outside of the fracture gap; lower line: the same images of an area within the fracture gap

Fig. 2

Scout view (left) and respective scanning area within two reference lines. Overview of 16 2D grey scale images out of the 620 slices with ROIs. The ROIs between the first and last slice were calculated by routine interpolation (right)

Fig. 3

2D axial μCT (first line) grey value image (contour of the ROI is not intended to be drawn directly adjacent to the bone surface) of the original cortical bone, callus and air/bone marrow with corresponding fully automatically segmented images below (callus = red, original bone = white, marrow/air = black). Corresponding histograms (a–f) identify the three areas. In c and d, grey values higher than or equal to 150 were coloured red, those higher than or equal to 370 white, and below 150 black, representing correct segmentation. In the second line, the threshold is set to low for the original cortical bone (a) and callus (b), leading to an overestimation of the respective tissue, whereas in the bottom line, the threshold is set to high for the original cortical bone (e) and callus (f), leading to an underestimation of the respective tissue

Fig. 4

Coronal half-sliced (first line) and axial (second line) 3D μCT reconstruction (Two Thresholding Algorithm) of a multi-fragmented fracture: callus (blue and semitransparent) and cortical bone (grey) (a), isolated cortical bone (b) and isolated peri- and endosteal callus as grey image (c) and colour-coded (d) The coloured local thickness map/histogram illustrates the thickness distribution of the trabeculae in d Note the isolated fractions of trabecular-like struts thicker than 1 mm in diameter (yellow) and the missing outer periosteal callus of a specimen from the prednisolone group. All images represent the same specimen. Scale bar = 1 mm

Fig. 5

Scatterplots with regression line showing the association between biomechanics (Load) and μCT (BV, TMD, BMC, SMI, DA, BS and Tb. Th.). The estimated regression model and correlation coefficient (r) are also indicated

Fig. 6

Scatterplots with regression line demonstrating the association between biomechanics (Stiffness) and μCT (BV, TMD, BMC, SMI, DA, BS and Tb. Th.). The estimated regression model and correlation coefficient (r) are also indicated