Background & Summary

The field of application of digital bone models is broad. Three-dimensional (3D) reconstructions of bones are used in biomechanics, biomedical engineering and medical image processing for musculoskeletal modelling1,2, finite element analyses3, statistical shape modelling4,5,6 or 3D reconstruction from sparse imaging data, such as radiographs7,8 or EOS images9. 3D reconstructions of the bones are used in orthopedics, traumatology or radiology for the development of implants10,11,12,13,14, surgical instruments15,16 or procedures, for diagnosis and decision-making17,18, preoperative planning19,20 and navigational guidance during computer assisted surgery8,21, the evaluation of outcome22, surgery simulation23, surgical education and training24, especially in the context of personalized, patient-specific, customized or individualized medicine. The surgical guidance based on bone models can be virtual, augmented25 or mixed reality26, or 3D printed27,28. Further fields of application are anatomy and patient education29,30, morphometrics31 and anthropometry32,33, forensic anthropology34,35, ergonomics, usability and human factors engineering36, accident and injury analysis and prevention37.

However, to the best of the author’s knowledge, no open access database or repository of skeletal surface models of the full lower extremities exists. Therefore, the objective of this study was to provide access to consistent complete bone models of the pelvis and lower limbs of multiple subjects. The database is supposed to enable other researches to quickly develop, test and verify new methods, approaches, algorithms or proofs of concept without the time-consuming and labor-intensive work of data collection and curation, segmentation and reconstruction. The database is expected to help the scientific community to facilitate research and improve the reproducibility and comparability of studies by giving access to the raw medical imaging data, including the metadata of the subjects and the segmentations and surface models of the bones. Hence, different researchers and research groups can resort to the same datasets for the validation of methods and comparison of results. Different deep learning models for artificial intelligence-based bone reconstruction, for instance, could be benchmarked by applying them to the raw computed tomography (CT) data and comparing the automatic with the manual segmentations of the database. The database can also be used as additional training data for existing deep learning models38,39.

Methods

Source of the raw CT data

The segmentations and models of the bones of the lower extremities were created from anonymized postmortem CT scans of the whole body originally published by Kistler et al. in the Swiss Institute for Computer Assisted Surgery Medical Image Repository (smir.ch) as open access Virtual Skeleton Database (VSD)40. The CT datasets were provided by the forensic institutes of the universities of Bern and Zürich and shared under the Creative Commons Attribution-NonCommercial-ShareAlike (CC BY-NC-SA) license after ethical approval of the Cantonal Ethics Committee Bern41. Further information about the datasets can be found in the literature cited40,41. Due to ongoing difficulties in accessing the SMIR website, the author decided to reupload the original datasets without any changes to the open access hosting service Zenodo: https://doi.org/10.5281/zenodo.827036442.

CAUTION

The VSD contains a few inconsistencies, such as duplicate CT datasets. The author of this publication is not connected to the SMIR or VSD and, therefore, not responsible for errors in the VSD. However, errors that the author recognized during the work with the VSD were logged and are reported in the reupload of the VSD42.

Subject selection

Twenty subjects (ten male and ten female) were selected from the VSD for the creation of the bone models with the objective of covering a wide age range.

The inclusion criteria were:

  • Availability of age, body weight and body height.

  • Integrity and completeness of the lower body’s skeletal anatomy.

The exclusion criteria were:

  • Difference between the gender specified in the metadata and the biological sex visible in the CT data.

  • Presence of artificial joints or bone fractures.

The average age, weight and height of the twenty subjects were 52 ± 21 years, 70 ± 13 kg and 1.7 ± 0.1 m, respectively. An overview of the subjects is presented in Table 1. Some subjects were processed before the inclusion and exclusion criteria were defined. Ten of the subjects did not meet the criteria. These ten additional subjects are also published as part of the database since they still might be useful for some applications, but they are tagged by a comment in the database so they can be easily identified by the user (see Table 1).

Table 1 Twenty complete subjects of the database and ten additional incomplete or inconsistent subjects.
Full size table

Reconstruction of the osseous anatomy

The bone surfaces were semi-automatically reconstructed by thresholding (Fig. 1). Two hundred Hounsfield units43 were chosen as the lower threshold and the maximum Hounsfield unit value present in the volume data was selected as the upper threshold. Subsequently, a manual post-processing using the software 3D Slicer (slicer.org) with default smoothing settings was performed44. The bones were manually segmented at the joints if necessary. All joints were segmented. However, some segments contain multiple components as follows:

  • Sacrum including the coccyx (if not fused with the sacrum)

  • Hip bone (also called pelvic, innominate or coxal bone)

  • Femur

  • Patella

  • Tibia

  • Fibula

  • Talus

  • Calcaneus

  • Tarsals, including the cuboid, navicular and three cuneiforms

  • Metatarsals

  • Phalanges

Separate segments were created for the left and right leg. Some segments contain small sesamoid bones if present. This applies to the metatarsals for all subjects but, in some cases, also to other bones, such as the femurs.

After the segmentation, the bones were reconstructed by manually closing holes present in the outer surface. No gap closing, hole filling or wrapping algorithms were used. The reconstructed surface models were exported as mesh files in the Polygon File Format (PLY) and imported into MATLAB using a conservative decimation and remeshing procedure (Fig. 1). The Hausdorff distance between input and output mesh was limited to 0.05 mm for the decimator. The adaptive remesher permitted a maximum deviation of 0.05 mm from the input mesh with a minimum and maximum edge length of 0.5 and 100 mm, respectively. The decimator and remesher are plugins of the software OpenFlipper (openflipper.org)45.

Fig. 1
figure 1

Workflow of the creation of the lower body’s bony anatomy surface models.

Full size image

CAUTION

Each reconstruction of anatomical structures from medical images is subject to cumulative spatial errors arising from each step of the process chain. While the section “Technical Validation” should give an impression of the error that can be expected from the workflow described, users of the database should take into account the risk of larger reconstruction errors depending on the application intended.

The bone models of each subject can be visualized by running the MATLAB or Python examples. One subject is presented in Fig. 2. The 3D reconstructions were created by the author as a private side project between 2017 and 2022. Parts of the database containing fewer subjects and only the pelvis and femurs were published previously as part of other studies of the author46,47. This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

Fig. 2
figure 2

Surface models of the lower body’s osseous anatomy of subject 002.

Full size image

Analysis of the surface models stored as MAT files

The database was searched for duplicate subjects using a two-stage registration process. Each pelvis was transformed into an automatically detected pelvic coordinate system based on the anterior pelvic plane using the iterative tangential plane method46. Subsequently, the sacrum of each subject was registered to the sacra of all other subjects using a rigid iterative closest points algorithm. Lower outliers of the root mean square error between the two registered sacra were examined. One duplicate subject was identified, excluded from the database and replaced by another subject.

Each bone model of all subjects was visually reviewed for internal cavities connected to the outer surface or connections between the inner and outer surface, and corrections were performed if necessary. The mesh topology was checked for the following errors using MATLAB:

  • Duplicate, non-manifold and unreferenced vertices.

  • Boundary, non-manifold and conflictingly oriented edges.

  • Duplicate and degenerated faces.

  • Self-intersections and intersections with adjacent bones.

The errors were corrected if present.

The volume enclosed by the outer surface of the bone models was calculated and is presented in Table 2. The values were compared with those from literature. However, caution must be applied since different definitions and measurement methods of the bone volume exist. Studies reporting the trabecular or cortical volume of the bones were not considered. The values of the bone volume correspond to those observed in previous studies48,49,50,51.

Table 2 Volume enclosed by the outer surface of the bone models of the twenty complete subjects of Table 1.
Full size table

Data Records

As mentioned above, a mirror of the complete VSD as hosted originally by Kistler et al. at smir.ch is available at Zenodo: https://doi.org/10.5281/zenodo.827036442.

The CT volume data, segmentations, reconstructions and raw PLY mesh files of the subjects of Table 1 are accessible via Zenodo: https://doi.org/10.5281/zenodo.830244852. The files of each subject are linked by a project file, called MRML scene file, that can be opened with the open-source medical imaging software 3D Slicer (slicer.org).

The post-processed mesh files of the subjects of Table 1 are stored as MATLAB MAT files, released as Git repository at https://github.com/MCM-Fischer/VSDFullBodyBoneModels and versioned via Zenodo: https://doi.org/10.5281/zenodo.831673053. The use of the MAT files is explained by examples for MATLAB and Python in the Git repository.

Technical Validation

The VSD also contains CT data of the European Spine Phantom that was introduced by Kalender et al. in 199554. The CT phantom data was used to evaluate the reconstruction process described above. After the creation of the surface model of the phantom, landmarks and areas were manually selected on the surface model of the phantom. Planes or cylinders were fitted to the areas selected to calculate the geometric parameters of the phantom. The errors between the reconstructed and the reference values of the geometric parameters reported in the publication by Kalender et al. are presented in Table 3. The mean error was 0.2 ± 0.4 mm and the mean absolute error was 0.4 ± 0.2 mm. This agrees well with accuracies reported in literature for 3D bone reconstruction using CT. Lalone et al. reported a mean error of 0.4 ± 0.3 mm for the cortical bone of the upper extremities55, Wang et al. reported a mean error of 0.5 ± 0.2 mm for machined bone specimens from the femur and tibia56 and van den Broeck et al. reported a mean absolute error of 0.5 ± 0.2 mm for the tibia57.

Table 3 Differences between the reconstructed values and the reference values of the geometric parameters of the European Spine Phantom54.
Full size table