May 17, 2022

findmeguilty-themovie

Technology Forever

Synthetic skull bone defects for automatic patient-specific craniofacial implant design

This study was approved by the ethics committee of the Medical University of Graz (MedUni Graz), Austria (EK-30–340 ex 17/18, Medical University of Graz, Austria), trial registration number DRKS00014853. Informed consent have been obtained for this experiment.

Figure 1 shows an overview of the workflow for data collection and preparation. First, clinical CT scans were retrospectively acquired. Second, the head was cropped from the scan. Third, the skull was extracted semi-manually from the CT scans using thresholding and case-specific post-processing strategies. Last, we converted the binary skull map to a 3D model in STL format.

Fig. 1

Data processing workflow illustrated using Case02. (a) original CT scan; (b) CT cropping; (c) thresholding (HU value ≥ 150) and CT table removal; (d) STL conversion.

Data processing workflow illustrated using Case02. (a) original CT scan; (b) CT cropping; (c) thresholding (HU value ≥150) and CT table removal; (d) STL conversion.

Data acquisition and selection

24 skull CT datasets have been collected retrospectively from May 2018 to August 2019 at the Medical University of Graz. CT data were generated according to a standard protocol, using a peak kilovoltage of 120 kVP. Amongst others a Siemens CT scanner (Sensation 64) was used for data generation. The data were originally provided in Digital Imaging and Communications in Medicine (DICOM) format. Only high-resolution CT scans with a voxel resolution of 512 × 512 × Z voxels, where Z ranges from 47 to 490, were selected for our collection. For this work, the original DICOM scans were anonymized by converting them to the Nearly Raw Raster Data (NRRD) format, where all DICOM tags, like patient name, age, sex, etc. are completely removed.

Skull segmentation

The segmentation of the skulls provided in our dataset were performed semi-manually by experts from the Medical University of Graz using the open-source software 3D Slicer (https://www.slicer.org)25.

Skull cropping

As a first step, the datasets, which originally covered varying regions of the patient, such as the whole torso or the head and neck area, were cropped to contain the region of the skull only. During cropping, it was made sure to retain all important structures required for cranial implant design, in particular the neurocranium. Most parts of the maxilla and mandible were excluded. An example for original and cropped CT scan is shown in Fig. 1(a,b), respectively.

Threshold-based segmentation

Even if the skull can be segmented from CT scans using simple thresholding, it is an experience-dependent task to decide for each CT scan the proper bone threshold, which also requires clinical knowledge in radiologic anatomy. The suitable threshold for each scan was defined in Hounsfield units (HU), which provide the density information of the bone. These can vary greatly with the age of the patient and possible comorbidities (e.g. osteoporosis). The density of calcium deposits must also be taken into account. An overview of the applied values for each case is given in tab:data. The result of this step is a binary map of the patient’s cranium, where a label of one corresponds to structures belonging to the skull, and a label of zero denotes the background.

Skull cleaning

Parts of the CT table are sometimes included in the CT scans, which we consider as noise, because it does not belong to the skull bone. Furthermore, in some cases, after thresholding a considerable amount of noise is left inside the skull, which results from the presence of some high density matter in the brain. During thresholding, both the skull and the noise are categorized as foreground, since they have a similar density and, therefore, similar HU values. The high density matter is a calcification within the soft tissue and a reflection of inlays. This is a result of high levels of calcium in the blood (hypercalcaemia), conditioned by bad lifestyle. During an infection or as part of the aging process, calcifications deposit in the soft tissue, including the brain. However, such calcium deposits or calcifications exist in the whole body. The resultant noise can be partly removed by setting the lower bound of the threshold higher, but then also parts of the skull with lower bone density will be removed. Therefore, noise, such as the CT table or speckles from calcifications within the skull were removed by only keeping the largest connected component from the threshold-based segmentation.

Aforementioned calcifications can also lead to uneven surfaces of the segmented skull when they are partly connected to it. Therefore, median smoothing with a radius of 2 mm was applied to those cases. The threshold-based segmentation can also lead to artifacts aside from noise, such as holes in the segmented surface in areas of very delicate, thin bony structures. In case of holes in the threshold-based segmentation, they were filled manually in a slice-by-slice basis. Other artifacts, such as streaking from strongly scattering dental materials, were also manually removed. tab:data shows the applied post-processing strategies for every case. A final, cleaned skull segmentation is visualized in Fig. 1(c).

3D model creation

Triangular meshes and point clouds are the basic data structures in computer graphics. To facilitate the usage of our skull datasets for researchers from the computer graphics field, we also provide 3D triangular meshes of the 24 skulls in the Stereolithography (STL) format with our data collection. The STL mesh is created from the aforementioned binary segmentation map using the open-source software 3D Slicer. After the binary skull map is generated, we used the’export to model’ functionality under the’Segmentations’ module of 3D Slicer to create the STL files with default settings. An example of the resultant mesh is shown in Fig. 1(d). Based on the marching cube algorithm, we also provided a python script on Github (https://github.com/Jianningli/SciData) to automatically convert binary skull label maps (in NRRD format) into meshes (in STL format).

Artificial defect injection

One of the aims of our data collection is to provide skulls without cranial defects to establish an atlas for cranial defect reconstruction. However, our collection can also be used as training and test set for algorithms reconstructing craniofacial defects. Therefore, we provide software scripts for the injection of artificial cranial defects into the healthy skulls within our data collection. The complete skulls and the skulls with artificially injected defects can be seen as equivalent to skulls obtained from pre-operative CT scans and post-operative CT scans of patients who had to undergo craniotomy. Figure 2(a,b) shows a healthy skull in 3D (A) and in 2D sagittal view (B). (D-E) show the corresponding defective skull in 3D (D) and 2D sagittal view (E). (F) shows the implant i.e., the portion removed from the healthy skull. (C) shows how the implant should match with the defected region on the defective skull in terms of bone thickness, boundary and shape.

Fig. 2
figure2

Illustration of defect injection to a healthy skull (Case02). A healthy skull in 3D (a) and 2D sagittal view (b). The corresponding defective skull in 3D (d) and 2D sagittal view (e). A portion of the skull (shown in gray) is removed (c). The removed portion, i.e., the implant in 3D (f).

Artificial defect

The artificial defects in the dataset are realistic but simplified compared to the real surgical defects. The shape, position and size of the real surgical defects do not have a fixed pattern and are determined by the pathological conditions e.g., the size and location of the brain tumor, of each specific patient. However, the real surgical defects do have something in common, i.e., intra-operatively an cranial drill (craniotome) is used in order to open the cranium of the patient and this course of action can leave a small roundish drilling gap on the boarder between the final implant and the defect (clinical implant offset), which does clinically not influence the bone healing process. On the one hand, the virtually created defects are realistic as we mimic the drilling process and also include a clinical implant offset, as it can be seen in Fig. 2(d). Additionally, the defect position is randomized and the extent of the defect size is varying. These characteristics of the virtually created artificial defects are also consistent with the real clinical situation and therefore comparably with the clinical practice. On the other hand, the artificial defects are simplified compared to real surgical defects. The border wall of real surgical defects tends to be rough and irregular, due to pathological or surgical reasons. However, the automatically injected defects, as it can be seen in Fig. 2(c), usually have smooth and ‘straight’ border walls.

Dataset enlargement

We create 10 random defects for each skulls. This process also artificially enlarges our dataset, making it applicable to training and evaluation of deep learning algorithms for automatic cranial reconstruction and implant generation. Figure 3 shows the ten defective skulls created out of Case02. It should be noted that even if the artificial defect presented in Fig. 2 does not exactly resemble the defects in a real craniofacial surgery, it demonstrates the feasibility of injecting realistic defects for future researchers using these data. Currently, we provide the python scripts for automatic injection of varied defects (e.g., roundish, triangular, rectangular, squarish and etc.) to the complete skull on Github (https://github.com/Jianningli/SciData).

Fig. 3
figure3

Illustration of ten different defects for Case02.

However, there are limitations, when generating a large number of artificial skull defects out of a few without bone defect. This course of action can increase the defect variations but the variations of the skull shape are limited to the original 24 skull bones. If used for training Deep Learning algorithms, the network cannot potentially generalize well to varied skull shapes. Creating multiple defects that are varied in terms of shape, size and position on each skull can increase the defect variations, which therefore can help deep learning algorithms generalize well to varied skull defects. There is currently no strict theoretical nor experimental findings about a quantitative threshold for the enlargement process. However, based on our experience, 10 randomly generated defects tend to be sufficient for deep learning algorithms to learn the defect variations26. Generating an excessive high amount of defects, e.g., hundreds or even thousands of defects per skull, is neither efficient nor necessary as excessive defect generation will produce redundant information needed to learn the defect variations. However, as the shape variations of the skull remains unchanged, the deep learning algorithms tend to be overfitting to the 24 skull shapes and can not generalize well on new skull datasets. This can only be overcome by increasing the amount of original cases in a future work. With this work we also hope to inspire other researchers to provide their cases to the research community. In doing so, these additional cases would also cover (different) scanners and scanning protocols from other institutions, thus increasing the variety and making deep learning algorithms more robust.

Defect injection toolbox

The current toolbox provided on Github (https://github.com/Jianningli/SciData) uses a cubic and spherical mask to ‘erase’ the skull bone that is overlapped with the mask. For each defect generation, the mask moves randomly within the skull to create randomized defects. Depending on where the mask is, the shapes of the resultant defect are not restricted to cubic and spherical, as can be seen from Fig. 3. The real surgical defects tends to be more irregular as it is specific to the pathological condition e.g., the position and size of the brain tumor of the patient. Usually, neurosurgeons use a cranial drill to open the cranium, resulting in a small roundish hole on the corner of the defect. Therefore, the toolbox offers, besides the normal cubic and spherical masks, more realistic defect masks to generate such small roundish holes on the defect corners, as can be seen from Figs. 2 and 3. The current toolbox can be used as a basis and can be extended to generate arbitrarily shaped, positioned and sized skull defects.