The work, published in Frontiers in Medical Engineering by Peter T. Shyu, Samuel T. Robinson, and X. Edward Guo, builds on the lab’s previous demonstration that time-lapse microCT imaging can quantify modeling and remodeling across bone microstructure. The new paper is a technical deep dive into the method of voxel-tracking, a computational approach that examines the sequence of bone events at each three-dimensional coordinate over an experimental timeline. The distinction the technique draws is fundamental to bone biology. Remodeling is bone formation and resorption coupled in space and time, carried out by basic multicellular units in which osteoclasts resorb bone and leave behind a scalloped surface, followed by a reversal phase that allows osteoblasts to form bone at the previously resorbed locations. Modeling, by contrast, is uncoupled: independent formation is called anabolic modeling, while independent resorption is called catabolic modeling. Whether these processes are coupled or uncoupled determines how bone adapts, and in metabolic bone diseases like osteoporosis the balance tips toward resorption, degrading bone mass, architecture, and strength.

Traditional methods for studying these dynamics have relied on 2D dynamic histomorphometry, in which multiple fluorochromes that chelate to calcium in mineralizing bone are administered at separated time intervals, while bone resorption is measured only indirectly. The approach comes with well-known technical difficulties: subjectivity in distinguishing smooth and scalloped bone cement lines, overlapping spectra from multiple fluorochrome labels, lack of labels compatible with osteoporosis therapeutics, and the need for ex vivo processing of 2D samples. Computed tomography has changed the landscape. Using image registration, the accurate and longitudinal identification of bone formation and resorption was first shown in the rat tibia in 2004, and the technique was subsequently developed to quantify dynamics in mouse tail vertebral trabecular and cortical bone and tibial cortical bone. But few studies have specifically quantified remodeling as bone formation following bone resorption at a specific location. The Columbia group’s method is designed to do exactly that, directly assessing longitudinal 3D modeling and remodeling across the bone microstructure.

The experimental system is elegant in its precision. The researchers applied mechanical loading to 16-week-old mice 5 consecutive days per week for 5 weeks, loading the right tibia by axial compression while the left tibia served as an internal, nonloaded control. A standard load frame applied a preload of 1 N compression for 4 minutes, followed by a haversine waveform between 1 N and 9 N compression for 100 cycles at 2 Hz. Critically, in vivo microCT imaging of the tibia began 2 weeks before the start of mechanical loading, with weekly follow-up imaging, so that the baseline state of the bone was fully characterized before any loading stimulus was applied. Scanning parameters included 55 kVp energy, 145 µA intensity, 300 ms integration time, and no frame averaging, with images reconstructed at 5 µm isotropic voxel size. Each tibia scan covered a 3.84 mm region in length, starting from the proximal growth plate and extending distally, with a scan time of 13.3 minutes per tibia under isoflurane anesthesia.

The computational heart of the method is the image registration and voxel-tracking pipeline. Full-resolution 16-bit grayscale images of extracted trabecular and cortical bone regions were registered and transformed to a common 3D position using Scanco Image Processing Language algorithms. To minimize bias from interpolation of voxel values during image transformation, the first pair of weekly microCT images were transformed to their midpoint, so that every weekly image is subject to similar interpolation effects. Images were first aligned by center of mass, followed by 3D rigid registration using a downhill simplex method to maximize the normalized correlation coefficient, with registration performed at three downscaling levels to avoid local maxima. Because the trabecular bone shifts distally over time due to the continually active growth plate in mice of this age, a two-step registration scheme was used: first a mask including the whole bone, followed by a trabecular bone-specific mask. Cortical bone regions required only one step using a mask of the entire cortical bone.

After registration, a Gaussian filter was applied to both reference and follow-up images to reduce noise, and images were segmented by applying global threshold values of 360 and 400 per mille for trabecular and cortical bone, respectively. The reference image was then subtracted from the follow-up image to identify bone formed and bone resorbed, which were saved as separate images. To capture bone formation at its earliest stage, as osteoid is first secreted and then mineralized over time, voxels with values greater than 75% of the global threshold value after image subtraction were also classified as formed bone. The resulting images were further filtered by removing clusters of less than 15 voxels, considered likely to be noise. Masking algorithms ensured that voxels classified as bone formed were labeled as bone in the follow-up image and as background in the reference image, and vice versa for voxels classified as bone resorbed. The segmented reference and follow-up images were added together to identify voxels labeled as bone in both time points, resulting in a constant bone image, and the three images were concatenated to generate weekly transition images with each voxel labeled with the bone event that occurred.

Quantifying bone volume changes was done by tracking the events at the voxel-level over the experimental timeline. Cut images were converted to text files listing each voxel’s coordinate and a value designating its bone adaptation event — formation, resorption, constant bone, or background — using a custom C-based program, and a custom MATLAB script was used to quantify bone volume changes at each voxel. The sequence of bone events at each voxel was characterized by applying a moving window to count the bone events at that voxel. With 5 weeks of mechanical loading, catabolic modeling can be quantified at the first 3 time points because week 3 is the last week with 2 follow-up images available to distinguish catabolic modeling. Bone volume changes, formation, and resorption events were normalized to their respective values at baseline prior to the start of loading, while modeling and remodeling values were normalized to the baseline bone volume.

The results reveal striking site-specific differences in how bone responds to mechanical loading. In the trabecular bone of control limbs, BV/TV decreased more drastically over time in wildtype mice due to greater levels of bone resorption and a decline in bone formation compared to SOST knockout mice. Loading increased wildtype BV/TV relative to control limbs due to elevated levels of bone formation, while the SOST knockout response was more attenuated and driven by a slight trend of decreased bone resorption. After further classification of formation and resorption into modeling and remodeling events, loading was observed to stimulate anabolic modeling and remodeling and reduce catabolic modeling throughout the course of loading in wildtype mice. The accumulated bone volumes by anabolic modeling and remodeling in wildtype mice increased after loading, and these values were greater than in SOST knockout mice. Catabolic modeling bone volume changes were decreased after loading in wildtype mice, particularly due to an early response in week 1, while the response in SOST knockout mice was more attenuated.

Cortical bone told a more nuanced story. Periosteal bone volumes increased more robustly due to loading, resulting in increases in overall cortical bone volume to levels above baseline for both genotypes. Endosteal bone volumes remained relatively stable over the course of loading, with a trend for an early increase in wildtype endosteal bone volumes that then decreased by week 5. Cortical formation peaked by week 2 for wildtype mice and then decreased with continued loading, while SOST knockout mice showed a more level formation response. The overall cortical bone response was dominated by modeling rather than remodeling. Loading affected wildtype modeling and remodeling from the onset of loading, particularly with an early increase in remodeling through week 2, and SOST knockout mice showed a similar but reduced response in accumulated changes of modeled and remodeled bone. In control SOST knockout limbs, there was an overall increase in modeling and remodeling compared to wildtype bone over time.

The implications reach beyond the mouse tibia. Sclerostin, encoded by the SOST gene, is the target of romosozumab, a humanized antibody approved for osteoporosis treatment with time-dependent effects on trabecular and cortical bone. Understanding how SOST depletion alters the coupling of formation and resorption under mechanical load provides mechanistic insight into how such therapies interact with exercise, a promising noninvasive treatment option for improving bone strength. The technique’s versatility extends to clinical imaging: time-lapse HR-pQCT with an effective dose of 3–4 µSv is increasingly used to evaluate human bone turnover, and the voxel-tracking approach could be adapted to suit different subjects and regions of interest. The authors note that longitudinal real-time subject-specific finite element analysis can be incorporated and combined with other segmentation techniques to label the trabecular microarchitecture by individual plates and rods. As the global burden of fractures continues to grow — affecting an estimated 1 in 2 women and 1 in 5 men over age 50 — techniques that directly assess modeling and remodeling dynamics in space and time will provide critical inputs for computational models of bone adaptation and disease progression, and function as a valuable tool for analyzing the long-term effects of therapies for bone diseases.

Subject of Research: In Vivo quantification of 4D modeling and remodeling in trabecular and cortical bone microstructure

Article Title: In Vivo quantification of 4D modeling and remodeling in trabecular and cortical bone microstructure

Article References: Shyu, P. T., Robinson, S. T., & Guo, X. E. (2025). In Vivo quantification of 4D modeling and remodeling in trabecular and cortical bone microstructure. Frontiers in Medical Engineering, 3, Article 1547895. https://doi.org/10.3389/fmede.2025.1547895

Image Credits: AI Generated

DOI: 10.3389/fmede.2025.1547895

Keywords: Vivo, quantification, modeling, remodeling, trabecular, cortical, bone, microstructure, scientific research