A new method for segmentation and analysis of bone callus in rodent fracture models using micro‐CT

Abstract Fracture burden has created a need to better understand bone repair processes under different pathophysiological states. Evaluation of structural and material properties of the mineralized callus, which is integral to restoring biomechanical stability is, therefore, vital. Microcomputed tomography (micro‐CT) can facilitate noninvasive imaging of fracture repair, however, current methods for callus segmentation are only semiautomated, restricted to defined regions, time/labor intensive, and prone to user variation. Herein, we share a new automatic method for segmenting callus in micro‐CT tomograms that will allow for objective, quantitative analysis of the bone fracture microarchitecture. Fractured and nonfractured mouse femurs were scanned and processed by both manual and automated segmentation of fracture callus from cortical bone after which microarchitectural parameters were analyzed. All segmentation and analysis steps were performed using CTAn (Bruker) with automatic segmentation performed using the software's image‐processing plugins. Results showed automatic segmentation reliably and consistently segmented callus from cortical bone, demonstrating good agreement with manual methods with low bias: tissue volume (TV): −0.320 mm3, bone volume (BV): 0.0358 mm3, and bone volume/tissue volume (BV/TV): −3.52%, and was faster and eliminated user‐bias and variation. Method scalability and translatability across rodent models were verified in scans of fractured rat femora showing good agreement with manual methods with low bias: TV: −3.654 mm3, BV: 0.830 mm3, and BV/TV: 7.81%. Together, these data validate a new automated method for segmentation of callus and cortical bone in micro‐CT tomograms that we share as a fast, reliable, and less user‐dependent tool for application to study bone callus in fracture, and potentially elsewhere.


| INTRODUCTION
Formation of mineralized callus is an integral part of indirect fracture repair, providing the bone with biomechanical stability, restoring limb function, and enabling later bone remodeling. The fracture callus is therefore a highly dynamic environment with changing proportions of tissues with varying architectural and material properties, which together determine the robustness of the repair tissue and ultimately the progress of bone regeneration. An appreciation of both the structural and material properties of the callus and repaired bone by appropriate analyses is therefore vital in developing our understanding of the fracture repair processes under different pathophysiological states and experimental settings. 1 Effective assessment of fracture repair requires evaluation of strength and stiffness of the fracture callus and repaired bone within the integrated context of whole bone structure. However, these mechanical parameters are currently difficult to measure directly in clinical settings. 2 Current clinical fracture assessment predominantly involves use of plain film X-ray imaging, 3 relying on two-dimensional (2D) images to assess 3D structures. The noninvasive, nondestructive method of computed tomography (CT) imaging can produce an array of 3D evaluations of structure and measurements of local mineralization. These can be used to indirectly estimate predicted strength and stiffness of the fracture callus and repaired bone. In a clinical research setting, quantitative computed tomography (QCT) or peripheral QCT (pQCT) have been used to assess bone mineral density (BMD) during fracture repair. 4 These studies have established that callus size, bone mineral content (BMC), and BMD from pQCT are useful surrogate measures of compressive failure load. 5 Micro-computed tomography (micro-CT) provides far greater resolution than clinical CT scanners and is ideal for studying the architecture of woven (irregular) bone in callus, particularly in the context of small animals. Several studies have successfully utilized micro-CT to measure an array of parameters associated with the architecture and mineralization of developing callus, with Morgan et al. 2 identifying bone volume (BV), bone volume/ tissue volume (BV/TV), and tissue mineral density as the strongest predictors of fracture callus mechanical strength. 2,6,7 In a research setting, micro-CT analysis provides a wealth of data and insight into the structure and mineralization of both repairing cortical bone and callus. However, current methods for appropriate segmentation of callus from cortical bone rely on semiautomated methods applied to defined regions of interest (ROIs). 2,[8][9][10][11] Defining these regions requires users to manually specify the boundary of the callus-where it meets endosteal and periosteal surfaces of the original bone-in multiple 2D images which makes this process both time-consuming and prone to operator error. Inappropriate segmentation results in inaccurate insights into healing progression and mechanical properties of the repairing bone. A novel automated method that addresses some of these limitations has been described. 12 The segmentation method that this elegant study employed is, nonetheless, reliant upon multilevel global thresholding and uses a single thresholding step with a view to define callus based on mineral density alone, which has limitations. These limitations may indeed prove problematic, as the callus increases in mineral density and radio-opacity during healing, making its discrimination from original or repaired bone highly subjective when reliant upon greyscale thresholding.
The need for quantitative evaluation of the fracture healing in an experimental setting, coupled with the lack of a ready method for selective and fast segmentation of mineralized callus, establishes a bottleneck in scope to monitor, understand and improve fracture healing processes, which may inform clinical therapies. Herein, we describe a newly developed, fully automated method to segment and separately analyse the callus and cortical bone from micro-CT images.
This method provides innovation as, unlike previous automated approaches, it is neither limited by the need for user ROI contouring nor by any exclusive reliance upon thresholding, to facilitate discrimination between callus and cortical bone. Five 11-12-week-old male C57BL/6 mice were anesthetized with 4% isofluorane (4 L/min oxygen) and maintained with 1.5%-2% isofluorane (2 L/min oxygen) throughout surgery. After induction of anesthesia, buprenorphine was administered subcutaneously (100 µl, 0.3 mg/ml, Vetergesic; Ceva Animal Health); the right thigh was shaved and cleaned using iodine. Femoral fractures were generated using a model devised and described by Zwingenberger et al., 13 in which postfracture stabilization was achieved using application of an external fixator 14 ( Figure 1). Postsurgery X-rays were performed under anesthesia to ensure correct fixator positioning. Mice were euthanized 6 weeks postsurgery using CO 2 and femora dissected and fixed for 48 h in 4% PFA before long-term storage in 70% ethanol.
Fixed samples had the external fixator removed and surrounding tissue around the bone removed.
Six female, 10-12-week-old, Wistar rat femurs were subjected to a well-established fracture healing model using an open femoral osteotomy with a 1.5 mm fracture gap. 15 They were anesthetized with 4% isofluorane (4 L/min oxygen) and maintained with 1.5%-2% isofluorane (2 L/min oxygen) throughout surgery. After induction of anesthesia, buprenorphine was administered subcutaneously (100 µl, 0.3 mg/ml, Vetergesic; Ceva Animal Health). Briefly, the rats had a modified Harrison type 1a external fixator applied on the craniolateral aspect of the left femora 16 after surgical osteotomy and distraction to the required gap size. With appropriate analgesia, the rats had unrestricted exercise for 5 weeks, before being euthanized.
Femora were retrieved and fixed in 10% buffered formaldehyde for 3 days before scanning.

| Scanning
Samples were wrapped in plastic film before scanning to prevent drying and scanned using a Skyscan 1172F (Bruker). Mouse femora were scanned with X-ray settings 50 kV and 200 µA, using an aluminum 0.5 mm filter and exposure time 960 ms using a pixel size of 5 µm. Rat femora were scanned with X-ray settings of 60 kV and Before any segmentations, volumes of interest (VOIs) were selected within CTAn 1. 18.4 (Bruker). In mouse femora, the fracture was identified and was used as a reference point for the selection of a 10% portion of the femur length extending both proximally and distally as the VOI (Figure 2A). A volume of 10% of femur length was selected to include the entire callus region but ensuring not to include regions that contained callus formation due to the external fixator pins. As the left femur had no callus the equivalent region was selected as a VOI. In rat femora, the presence of the external fixator during scanning caused X-ray diffraction artifacts at sites in close proximity, therefore selection of a VOI required the center of the fracture gap to be identified as a reference point for the selection of 300 slices extending both proximally and distally (total of 600 slices,  5-μm thick slices) as the VOI. After selection, VOIs underwent segmentation, both manually and using the automated method (as described below) to separate callus from cortex ( Figure 2B). This process was completed using CTAn 1.18.4 (Bruker) but the outlined steps can readily be performed in any software with segmentation capabilities (e.g., ImageJ). Volume-rendered 3D visualizations were created using CTVox 3.3 ( Figure 2C-F) (Bruker). It is important to note that the automated segmentation method can be used with larger VOIs however the VOIs for this study were selected to ensure analysis of only fracture callus and not callus formation around the external fixator pins.

| Automated method
Using the selected VOIs, images were subjected to numerous 2D processing steps to achieve callus: cortex separation; this involved implementation of the preprogrammed image processing plugins within CTAn custom processing menu. The first step in this segmentation involves application of appropriate global thresholding (range 150-255 gray scale values) to separate "space" from bone, which simultaneously also removes the majority of the callus due to its lower mineral density. After thresholding, a 1-pixel radius round kernel "erosion" step that removes remnant callus is followed by a 3-pixel radius round kernel closing step, which replaces pixels lost within the cortex in the previous step. Ensuing rounds of despeckling to remove remnant callus are followed by morphological operations that replace pixels lost from the cortex during preceding erosion and despeckling. The image is then "despeckled" to remove connected white pixels (mineralized tissue) (area < 250 pixels) followed by a 2-pixel radius round kernel closing step.
Further "remnant" callus is removed with a 1-pixel radius round kernel erosion step and despeckling to remove connected white pixels (mineralized tissue) (area < 250 pixels); pixels lost from cortex are replaced with sequential closing steps consisting of 2 pixels, then a 10-pixel radius round kernel. Processing proceeds with a 1-pixel radius round kernel erosion step and two dilation steps of 4 and 2 pixels, respectively. A pair of final despeckling steps to remove connected white pixels (mineralized tissue) (area < 1000 and

| Manual segmentation
Using the selected VOIs, callus was manually segmented using CTAn (Bruker), first by segmenting the cortices using the manual ROI drawing tool to select the boundaries of the cortical bone and then using a tool called subtractive ROI to select everything outside of the manually drawn cortical ROI. This callus VOI and the cortical bone segmentations were then analyzed using 2D and 3D analyses within the custom processing tools in CTAn (Bruker).

| Validation
Validation of our method was achieved by comparing measurements of callus microarchitecture (tissue volume [TV], BV, and BV/TV) from the automated method with the "gold standard" manually segmented callus VOIs from scans of fractured femora. Nonfractured femora were also subjected to the same methodology in the absence of callus to demonstrate the accuracy in callus identification and separate segmentation from cortices.     Our data show that data sets segmented by different users only differed in measurement of cortical bone thickness. The interuser variation associated with manual segmentation between two different users in all parameters ranged from 1.72% for BV to 4.2% for connectivity ( Table 2).

| Nonbiased automatic segmentation is time efficient without user influence
To overcome inefficiencies related to manual segmentation a series of noise reductions, morphological, bitwise, and arithmetical operations were developed ( Figure 2B). Implementation of these methods shows that automatic segmentations of the same datasets used for the manual segmentation took~1 min per data set, a 30-fold reduction in time. Moreover, implementation of this automated methodology appears to restrict user variation.
Quantitative comparison of auto-segmented fractured and control/ nonfractured femora revealed that callus can be effectively segmented from cortical bone with proper and accurate selection. 3D volumetric analysis of the auto-segmented callus showed statistically significant differences between fractured and control (nonfractured) femora in all architectural parameters excluding BV/TV, with callus from fractured femora exhibiting higher BV, TV, BS, and connectivity ( Table 3). Comparison of 2D analytical profiles between automatic and manually segmented cortical datasets revealed a significant T A B L E 2 Interuser variation associated with manual segmentation Note: Ratio paired t-test was used to compare results of manual segmentation between Users 1 and 2. Data (means ± SD) with statistical significance as p ≤ 0.05 and CV (%) between users.
Abbreviation: CV, coefficient of variation. difference in B.Ar in the proximal callus region, this trend was also seen in the distal region but was not significant. However, there is no effect on shape, reflected by eccentricity in these areas as well as having no significant impact on mean polar moment of inertia and therefore predicted strength of the bone ( Figure 5A-D).

| Automatic segmentation of fracture callus from rat femora
To test the translatability of our automated segmentation method within rodent species we compared 3D analytical results from manual and automated segmented callus VOIs from rat femora. Using  (Table 4 and Figure 6A-C).
To further test the alliance between these methods we  segmentations in regions further away from the fracture site where the transition of callus to cortical bone is more advanced, making distinction increasingly difficult however these differences did not reach significant levels.
Analysis of manually segmented datasets found relatively small interuser variation of 1.7%, 3.7%, and 4.19% for BV, TV, and connectivity, respectively, there is no variation in the same datasets subjected to the automatic segmentation.
The contribution of cortical bone to the callus environment is often overlooked, however, we also demonstrated that this automatic method effectively segments cortical bone as a data set allowing for analysis. Cortical parameters showed lower variation between users compared with callus, however, there was a significant difference in cortical thickness. This was attributed to users finding it difficult to define the cortex and callus as separate ROIs, even within well-specified sample regions, or due to difficulties to visually identifying their specific boundaries ( Figure 2G). This highlights the subjective nature of manual segmentation and the variation introduced by users that is alleviated completely by our automatic segmentation method. However, significant differences were  segmentation of cortical bone becomes inaccurate presumably due to less distinction between cortical and callus bone in these most distant regions (see Figure 5C).
The past two decades have seen an increase in the use of rodent models to study fracture healing, which historically had been assessed in large animals like sheep, 18 dogs, 19 and pigs. 20 Rodent models are particularly useful to study fracture healing and the efficacy of treatment due to the ease and speed of breeding as well as the availability of genetically modified strains and molecular diagnostic tools. 21   Beyond these difficulties in efficient segmentation of particular callus areas, we must also consider the fracture model employed in our studies was not the more commonly used Einhorn model 22 but the model developed by Zwingenberger et al. 13 The former may allow more bone displacement or fragmentation and further validation of our automated method should therefore be considered before application in other rodent fracture models.
It is noteworthy that our studies segmented the callus at 5 weeks for rats and 6 weeks for mice, which are both later postfracture timepoints than those conducted using other automated methods previously (at 3 weeks). 12 This later timepoint of healing provides insight into fracture callus architecture beyond the timeline of current automated methods.
Other methods including our own 23 are described for semiautomated bone segmentation but have not been applied to callus. The excellent methods of Buie et al. 24 and Kohler et al. 25 solved many issues related to automated segmentation of cortical from trabecular bone.
These were designed for specific morphologies and required certain conditions about bone distribution to be met; relating to the diameter (in pixels) of the largest pore connecting the marrow to the exterior; the former also relied on dual thresholding. Our method addresses such limitations with a simple preprocessing step to isolate the region of interest, which enables all further steps to run automatically. It is important to emphasize, nonetheless, the conceptual similarity of these methods; indeed their independent development highlights their usefulness. Our method also protects against artefactual "clipping" of the outer cortical boundary that can arise in earlier methods if spaces outside the region of interest are smaller than the dilation and erosion amounts; the "cortical mask" grows beyond the image boundaries, while our method retains the periosteal boundary of the segmented cortex.
This does not negate the need to further validate our approach in future research via direct comparison with another well-described approach, such as histology in selected regions. Finally, it is also worth noting that neither the methods of Buie et al. 24 nor Kohler et al., 25 have been applied to the analysis of fracture callus.
In conclusion, we have defined a new automated method for segmentation of callus and cortical bone in micro-CT tomograms which is easy to use, consistent and comprehensive, allowing for bone microarchitecture analysis across entire fractures. It allows without a priori assumption regarding callus mineralization density predictions of mechanical strength to be made. In addition, as use of rodent models has become a standard in fracture studies, the translatability and scalability of our method in both mice and rats add to its potential utility.
It is envisioned that future development of invivo micro-CT machines may make it feasible to similarly assess fracture repair in humans to establish efficacy of new treatment strategies.

AUTHOR CONTRIBUTIONS
Mark Hopkinson: experimental design, experimental work, and manuscript preparation. Gareth Jones: experimental work and manuscript preparation. Lucinda Evans: experimental work.