Machine learning for automated 3-dimensional segmentation of the spine and suggested placement of pedicle screws based on intraoperative cone-beam computer tomography

Gustav Burström Department of Clinical Neuroscience, Karolinska Institutet;
Department of Neurosurgery, Karolinska University Hospital, Stockholm, Sweden;

Search for other papers by Gustav Burström in
jns
Google Scholar
PubMed
Close
 MD
,
Christian Buerger Digital Imaging, Philips Research, Hamburg, Germany;

Search for other papers by Christian Buerger in
jns
Google Scholar
PubMed
Close
 PhD
,
Jurgen Hoppenbrouwers Image Guided Interventional Therapy, Philips Healthcare, Best, The Netherlands; and

Search for other papers by Jurgen Hoppenbrouwers in
jns
Google Scholar
PubMed
Close
 MSc
,
Rami Nachabe Image Guided Interventional Therapy, Philips Healthcare, Best, The Netherlands; and

Search for other papers by Rami Nachabe in
jns
Google Scholar
PubMed
Close
 PhD
,
Cristian Lorenz Digital Imaging, Philips Research, Hamburg, Germany;

Search for other papers by Cristian Lorenz in
jns
Google Scholar
PubMed
Close
 PhD
,
Drazenko Babic Image Guided Interventional Therapy, Philips Healthcare, Best, The Netherlands; and

Search for other papers by Drazenko Babic in
jns
Google Scholar
PubMed
Close
 MD
,
Robert Homan Image Guided Interventional Therapy, Philips Healthcare, Best, The Netherlands; and

Search for other papers by Robert Homan in
jns
Google Scholar
PubMed
Close
 BSc
,
John M. Racadio Interventional Radiology, Cincinnati Children’s Hospital Medical Center, Cincinnati, Ohio

Search for other papers by John M. Racadio in
jns
Google Scholar
PubMed
Close
 MD
,
Michael Grass Digital Imaging, Philips Research, Hamburg, Germany;

Search for other papers by Michael Grass in
jns
Google Scholar
PubMed
Close
 PhD
,
Oscar Persson Department of Clinical Neuroscience, Karolinska Institutet;
Department of Neurosurgery, Karolinska University Hospital, Stockholm, Sweden;

Search for other papers by Oscar Persson in
jns
Google Scholar
PubMed
Close
 MD, PhD
,
Erik Edström Department of Clinical Neuroscience, Karolinska Institutet;
Department of Neurosurgery, Karolinska University Hospital, Stockholm, Sweden;

Search for other papers by Erik Edström in
jns
Google Scholar
PubMed
Close
 MD, PhD
, and
Adrian Elmi Terander Department of Clinical Neuroscience, Karolinska Institutet;
Department of Neurosurgery, Karolinska University Hospital, Stockholm, Sweden;

Search for other papers by Adrian Elmi Terander in
jns
Google Scholar
PubMed
Close
 MD, PhD
Full access

OBJECTIVE

The goal of this study was to develop and validate a system for automatic segmentation of the spine, pedicle identification, and screw path suggestion for use with an intraoperative 3D surgical navigation system.

METHODS

Cone-beam CT (CBCT) images of the spines of 21 cadavers were obtained. An automated model-based approach was used for segmentation. Using machine learning methodology, the algorithm was trained and validated on the image data sets. For measuring accuracy, surface area errors of the automatic segmentation were compared to the manually outlined reference surface on CBCT. To further test both technical and clinical accuracy, the algorithm was applied to a set of 20 clinical cases. The authors evaluated the system’s accuracy in pedicle identification by measuring the distance between the user-defined midpoint of each pedicle and the automatically segmented midpoint. Finally, 2 independent surgeons performed a qualitative evaluation of the segmentation to judge whether it was adequate to guide surgical navigation and whether it would have resulted in a clinically acceptable pedicle screw placement.

RESULTS

The clinically relevant pedicle identification and automatic pedicle screw planning accuracy was 86.1%. By excluding patients with severe spinal deformities (i.e., Cobb angle > 75° and severe spinal degeneration) and previous surgeries, a success rate of 95.4% was achieved. The mean time (± SD) for automatic segmentation and screw planning in 5 vertebrae was 11 ± 4 seconds.

CONCLUSIONS

The technology investigated has the potential to aid surgeons in navigational planning and improve surgical navigation workflow while maintaining patient safety.

ABBREVIATIONS

CBCT = cone-beam CT; IQR = interquartile range; MaxDist = maximum distance; OR = operating room; RMSDist = root-mean-squared distance.

OBJECTIVE

The goal of this study was to develop and validate a system for automatic segmentation of the spine, pedicle identification, and screw path suggestion for use with an intraoperative 3D surgical navigation system.

METHODS

Cone-beam CT (CBCT) images of the spines of 21 cadavers were obtained. An automated model-based approach was used for segmentation. Using machine learning methodology, the algorithm was trained and validated on the image data sets. For measuring accuracy, surface area errors of the automatic segmentation were compared to the manually outlined reference surface on CBCT. To further test both technical and clinical accuracy, the algorithm was applied to a set of 20 clinical cases. The authors evaluated the system’s accuracy in pedicle identification by measuring the distance between the user-defined midpoint of each pedicle and the automatically segmented midpoint. Finally, 2 independent surgeons performed a qualitative evaluation of the segmentation to judge whether it was adequate to guide surgical navigation and whether it would have resulted in a clinically acceptable pedicle screw placement.

RESULTS

The clinically relevant pedicle identification and automatic pedicle screw planning accuracy was 86.1%. By excluding patients with severe spinal deformities (i.e., Cobb angle > 75° and severe spinal degeneration) and previous surgeries, a success rate of 95.4% was achieved. The mean time (± SD) for automatic segmentation and screw planning in 5 vertebrae was 11 ± 4 seconds.

CONCLUSIONS

The technology investigated has the potential to aid surgeons in navigational planning and improve surgical navigation workflow while maintaining patient safety.

In Brief

The authors investigated whether a machine learning–based technology is able to identify vertebrae and their anatomical components and suggest optimal screw paths automatically for pedicle screws. Testing the system on a data set from 20 previously treated patients showed that the system suggested correct screw placement paths in 95% of the cases. The technology has the potential to improve workflow for surgeons and be a key step in automation of robotic assistance for appropriately selected cases.

Pedicle screw placement performed in spinal fixation procedures demands surgical expertise and poses risks for complications, as vital neurological and vascular structures are in close anatomical relation to the pedicles. The traditional method of using a freehand technique relies on anatomical landmarks, preoperative imaging, and use of x-ray fluoroscopy. Still, reported pedicle screw placement accuracies in a meta-analysis by Kosmopoulos and Schizas showed a weighted mean of 87.3% (median 79%) in the lumbar spine and 56% (median 94.3%) in the thoracic spine for procedures performed without navigation equipment.22 Another meta-analysis by Gelalis et al., using more stringent inclusion criteria, reported the percentage of screws with cortical violations > 4 mm to be 1%–6.5% for non-navigated screw placement and 0%–3.3% for navigated screw placement.13 This has led to an increasing interest in the use of surgical navigation systems to aid in pedicle screw placement.8,16 It has previously been shown that the use of surgical navigation coupled with intraoperative 3D imaging reduces the frequency of pedicle screw misplacements compared to the use of preoperative CT-based 3D navigation as well as 2D navigation and the freehand method.9,25,30,33,36 However, different variants of intraoperative 3D imaging have not been shown to significantly alter overall accuracy.32

With the increasing use of intraoperative 3D surgical navigation, valuable operating room (OR) time is spent on planning navigation trajectories,12 and perceived increased OR time is one of the main concerns cited when surgeons refrain from adopting the technology.15 Initial approaches to compensate for increased navigation time focused on automating identification of vertebral levels.27 More recent methods have focused on segmentation (building a virtual 3D reconstruction) of entire sections of the spine on preoperative CT image sets to aid in surgical planning in a fashion similar to segmentation systems used for cranial surgery.4,18–20,24 Commercially available navigation systems have recently introduced different forms of vertebral segmentation aids for intraoperative CT imaging to improve surgical navigation usability in the OR.14 However, to our knowledge, no published studies have validated the clinical accuracy of systems aimed at automatizing screw path planning.

In this study, we present and evaluate an automatic system (linked to Philips’ augmented reality surgical navigation) for intraoperative segmentation of vertebrae and identification of pedicles, including screw placement suggestions. The aim of the study was to develop and validate the segmentation system and determine whether it results in clinically relevant identification of pedicles for screw placement. Furthermore, we aimed to identify possible spine pathologies where the system might fail in correctly segmenting the vertebrae. Using machine learning methodology, we trained the segmentation algorithm on 21 human cadavers. We validated and determined the technical accuracy of the segmentation using a 3-fold crossover test on the same cadavers.3 Subsequently, we applied the system to segment each spine in a cohort of 20 clinical cases to evaluate the accuracy of pedicle identification. In addition, a qualitative evaluation of the segmentation output from the clinical cases was performed to judge whether it was adequate to guide surgical navigation or not. Two surgeons performed the evaluation of clinical relevance independently.

Methods

A System for Automatic Segmentation of Vertebrae, Identification of Pedicles, and Placement of Pedicle Screws

A cone-beam CT (CBCT) mounted on a robotic flat detector C-arm system (AlluraClarity, Philips Healthcare) was set up in a hybrid OR to mimic the intraoperative use of the automatic segmentation system in spinal surgery. The application of the system for intraoperative use is CE (Conformité Européene) marked but not FDA approved. Twenty-one cadavers were imaged from C1 to S5 using the CBCT. All cadavers were donated bodies, and their inclusion in the study was in accordance with the ethical permits of the local ethics committee. Each CBCT acquisition was used to generate volume data sets with an isotropic voxel size of 0.49 mm3 and a CT-like image quality. A total of 430 vertebrae were scanned and included in the data set. The complete cortical border of each vertebra was manually outlined (annotated) in 3D to serve as a reference for subsequent validation. A segmentation algorithm was developed that relies on a generic pre-generated 3D vertebral model and adapts it to the image. The generic 3D model is based on 3D-reconstructed scans of an anatomically correct plastic model of the spine. The segmentation algorithm follows a stepwise process outlined below and illustrated in Fig. 1.

  • Step 1: The algorithm requires information about which vertebral level is to be segmented. Once the vertebral identity is provided, a pre-generated 3D model of the corresponding vertebra is placed in the image at the correct vertebral level using a generalized Hough transform.1

  • Step 2: A parametric adaptation is applied to improve the alignment between the 3D model and image data set with respect to general features such as rotation, translation, and sizing.10,29,35

  • Step 3: The 3D model is locally deformed so that its boundaries align with the vertebral boundaries in the image. This fine-tunes the alignment between the 3D model and image with respect to local image features in a process called deformable adaptation.10,29

FIG. 1.
FIG. 1.

A visualization of how the software identifies the vertebrae by a 3-step process, showing the application of the generalized Hough transform (A), the parametric adaptation improving the global alignment (B), and the deformable adaptation finalizing the local alignment (C). Figure is available in color online only.

By comparing the final 3D reconstruction to the corresponding manually outlined vertebrae, the outcome of the segmentation could be evaluated.29 Using machine learning methodology, a segmentation algorithm (i.e., the 3 steps above) was trained and tested using a 3-fold cross-validation approach for each vertebral level. This approach meant that each segmentation model was trained on 66.6% of the cadavers and tested on the remaining 33.3%. Each vertebra was further divided into 6 parts (Table 1) for additional segmentation accuracy analysis. Technical accuracy of the final segmentation algorithm was measured by calculating the surface-to-surface distances between the 3D reconstruction and the corresponding manually outlined reference vertebra. Specifically, the root-mean-squared distance (RMSDist) and the maximum distance (MaxDist) were evaluated, as is common practice in the field of 3D image analysis.2,31

TABLE 1.

Mean distances between true vertebra surface and segmented model

RMSDist (mm)
Vertebral Levels*Entire VertebraeLeft PedicleRight PedicleSuperior EndplateInferior EndplateSpinous Process
Cervical (n = 451)0.810.520.670.870.901.30
Thoracic (n = 718)0.600.400.400.660.560.86
Lumbar (n = 254)0.810.600.530.690.880.88
All (n = 1423)0.700.470.510.730.731.00

RMSDist = root-mean-squared distance.

Each segmentation included 5 vertebrae. Consequently, some vertebrae within the CBCT were segmented more than once.

As shown in Fig. 2, five vertebral levels are automatically segmented at a time. This matches the number of levels typically included in a CBCT and also improves segmentation accuracy as more anatomical information is provided. Using the same approach, the system also subsegments specific anatomical parts of the vertebrae and, aided by mathematically defined criteria, is able to identify the pedicles and suggest an optimal screw path.

FIG. 2.
FIG. 2.

Visualization of the automatic pedicle identification and screw placement suggestion in axial (A) and sagittal (B) views and the visualization of the segmented vertebral section including screw suggestion (C). Figure is available in color online only.

Retrospective Application of the Automatic Segmentation System to Clinical Cases

To test the accuracy and usefulness of the system outlined above, the radiological data of 20 patients that had previously been enrolled in a clinical study using augmented reality surgical navigation were retrospectively analyzed.11 All patients had signed informed consent, and the retrospective analysis of the radiological data was approved by the local ethics committee. All patients had been scanned using the same type of flat detector C-arm system as was used in the initial cadaveric experiments. All thoracic and lumbar vertebrae that had been scanned corresponding to the surgical indication for each patient were segmented using the approach described above.

Pedicle Midpoint Accuracy in Clinical Cases

To quantify the accuracy of the clinically relevant part of the vertebral segmentation for placing pedicle screws in the thoracic and lumbar vertebrae, the midpoint of each pedicle was manually identified in the clinical CBCT scans on all vertebrae. The midpoint was defined as the middle of the pedicle in the coronal plane, at the narrowest part of the pedicle as seen in an axial view. The 2D distance from the manually identified midpoint of each pedicle to the midpoint of the corresponding 3D reconstruction of the pedicle (in the same coronal plane) was measured for all segmented pedicles as illustrated in Fig. 3.

FIG. 3.
FIG. 3.

Illustration of midpoint measurements. The distance (in red) from the midpoint of each pedicle (in brown) to the midpoint of the corresponding segmented 3D reconstruction of the pedicle (in blue) in the same coronal plane was measured for all segmented pedicles. Figure is available in color online only.

Evaluation of Segmentation Results in Clinical Cases

Using the clinical data sets, 2 independent reviewers (G.B. and O.P.) were instructed to evaluate each segmented pedicle according to the clinical usefulness of the segmentation.11 Each pedicle segmentation was rated as correct or incorrect by the 2 reviewers; a segmentation was defined as correct only if both reviewers judged that a screw placed solely on information from the 3D reconstruction would result in a correctly placed pedicle screw in the actual vertebra.

Statistical Analysis

A p value less than 0.05 was considered significant. Midpoint deviations and segmentation times were expressed as means ± standard deviations and as medians (presented with interquartile ranges [IQRs]). Surface-error distances in the form of root-mean-squared distances (presented with full ranges, minimum–maximum) were used in calculating errors between the surface of the 3D reconstructions and the corresponding manually outlined vertebral borders. A 2-sided Welch 2-sample t-test was used to evaluate mean differences between groups and confidence intervals. When reporting the evaluation of segmentation results in clinical cases only cases in which both reviewers judged a pedicle as correct were reported as such. However, for analyzing the interrater agreement, Cohen’s kappa was used.5 Statistical analysis was performed using RStudio (RStudio Team [2016]. RStudio: Integrated Development for R. RStudio, Inc.).

Results

Segmentation Surface-Error Distances on Cadavers

Each individual segmented vertebra was validated against the corresponding manually outlined reference vertebra for all spinal levels (C1–L5) in all cadavers. Overall, the segmentation error was a root-mean-squared distance (RMSDist) of 0.7 mm with a maximum distance (MaxDist) of 17.8 mm. The MaxDist of 17.2 represents an erroneous whole vertebra shift in the segmentation due to high local similarities between adjacent vertebrae. Segmentation errors varied between individual vertebral levels and between the cervical, thoracic, and lumbar regions (Table 1). The lowest segmentation error was at T1 (RMSDist 0.5 mm, MaxDist 3.7 mm) and the highest was at C6 (RMSDist 1.1 mm, MaxDist 17.2 mm). The mean time for loading image files into the software and executing a segmentation of 5 levels and suggesting screw paths was 11 ± 4 seconds.

Pedicle Segmentation Success on Patients

In total, 316 pedicles were segmented based on intraoperative CBCT images from 20 clinical cases. Among the segmented pedicles, 76 were in the lumbar spine and 240 in the thoracic spine. Spine pathology and associated pedicle segmentation accuracy are shown in Table 2. Overall, clinically adequate automatic pedicle segmentation was achieved in 86.1% of pedicles. This represents only pedicles where both reviewers judged the pedicle identification to be correct. Post hoc analysis of interrater agreement showed an absolute interrater agreement of 94.6% and a Cohen’s kappa of 0.73.5 The count of correctly and incorrectly segmented pedicles in each patient is seen in Fig. 4.

TABLE 2.

List of patients, their diagnoses, and the success rate of pedicle identification according to 2 separate reviewers

Patient No.DiagnosisTotal No. of Pedicles SegmentedNo. of Incorrect Segmentations% Correct Segmentations
1Spondylolisthesis40100.0
2Scoliosis240100.0
3Scoliosis16193.8
4Kyphosis22195.5
5Spondylolisthesis6266.7
6Scoliosis18194.4
7Degenerative disc40100.0
8Scoliosis22290.9
9Spondylolisthesis4250.0
10Stenosis6433.3
11Scoliosis240100.0
12Scoliosis220100.0
13Scoliosis22959.1
14Post-fracture kyphosis10280.0
15Scoliosis240100.0
16Scoliosis8187.5
17Scoliosis160100.0
18Scoliosis20195.0
19Scoliosis201145.0
20Scoliosis24770.8
Total3164486.1
FIG. 4.
FIG. 4.

The number of segmented pedicles in each patient separated by correct or incorrect identification. Patients with any of the exclusion criteria present are indicated by pound signs (#). Figure is available in color online only.

A further subanalysis was carried out to identify patient characteristics where the algorithm failed frequently. It was noted that the main number of failures occurred in patients either where the Cobb angle was > 75°, where previous surgery had severely changed the anatomy of the vertebrae, or where severely degenerated vertebrae were present. Patients belonging to any of these three categories accounted for 75% of the failed pedicle segmentations. By requiring manual pedicle identification by surgeons in these cases, a clinical accuracy of 95.4% was achieved in the remaining patients. Among these patients, there was no statistically significant difference in failure rate between lumbar and thoracic vertebrae or between individual vertebrae. The exclusion criteria used and associated failure rates are presented in Table 3.

TABLE 3.

Proposed exclusion criteria for patient characteristics that were associated with a high degree of failure (> 30%) when the software was tasked with identifying the pedicles

CriterionPedicle Segmentation Failure Rate
Cobb angle >75°34.8% (16/46)
Severe vertebral degeneration66.7% (4/6)
Previous surgery on vertebrae54.2% (13/24)

Values in parentheses are number of failures/total number of pedicles meeting the specified criterion.

Pedicle Midpoint Accuracy on Patients

The median distance between the 2D pedicle midpoint, measured manually on the CBCT image, and the midpoint measured on the corresponding 3D pedicle segmentation was 0.9 mm (range 0–18.3 mm). When excluding patients with the above-mentioned exclusion criteria (Table 3), the median midpoint error distance for the remaining pedicle segmentations was 0.8 mm (range 0.1–6.1 mm). The spread of midpoint distances can be seen in Fig. 5. When comparing midpoint error distances of correctly versus incorrectly identified pedicles, there was a statistically significant difference of 2.5 mm (p < 0.001; 95% CI 1.2–3.8 mm). The midpoint error distances were greater among pedicles judged to be incorrectly identified.

FIG. 5.
FIG. 5.

Distribution of distances between the algorithm-suggested pedicle midpoint and the actual manually defined pedicle midpoint. Deviation distances are stratified by correct or incorrect pedicle identification and median (solid vertical line) and quartiles (dashed vertical lines) are highlighted. Figure is available in color online only.

Discussion

Automatized and computer-assisted surgery is rapidly gaining ground.26,28 A driving force in this development is improved accuracy and the promise for better patient outcomes that follows.22,34 In addition, the speed and ease of use provided by available systems seems to be ever improving. Even though computers and robots may be able to perform certain tasks very accurately, there is yet no substitute for the surgeon’s intraoperative judgment. However, if supporting systems are flawed in providing correct data, so too may the surgeon’s decision making be flawed. Notably, only a few studies have assessed the clinical accuracy of automatic vertebral segmentation for intraoperative imaging and surgical navigation use. The present study is to the best of our knowledge the first to evaluate a method for intraoperative automatic vertebra segmentation, pedicle identification, and subsequent pedicle screw placement suggestion.

Since the cadaveric spines on which the segmentation algorithm was trained lacked severe spinal pathology, a follow-up analysis was performed on imaging data from clinical cases of spinal pathology. As the goal of segmentation was the placement of pedicle screws, we focused on this aspect of the segmentation. Overall a correct segmentation of pedicles was performed in 86.1% of the cases, and when our proposed exclusion criteria were applied a clinical accuracy of 95.4% was achieved (Table 3). However, the system requires that all suggested segmentations and screw paths are actively confirmed by the surgeon to further safeguard against negative outcomes for patients. Furthermore, although the system provides a suggested path, the choice of screw diameter and length is determined by the surgeon, based on the overall instrumentation construct and patient-specific circumstances. The accuracy reported here is the accuracy of correctly segmented pedicle borders, and not actual screw placement grading (e.g., Gertzbein or similar) as this would be dependent on several factors, including screw dimensions and adherence to the navigational plan. Future addition of ever more complex cases to the algorithm training material is expected to improve the accuracy in clinical cases.

As was shown in this study, a subset of patients appears to be more challenging for this segmentation algorithm. Cobb angles > 75°, extreme vertebral degeneration, and previous surgery with partial removal of vertebrae were all noted to lead to higher rates of failure, as could be expected. Previous studies have mainly reported success rates on either undefined or normal patient material.6,18,21 Lo et al., however, reported a very high success rate in identifying a correct vertebral level in a consecutive series of patients undergoing primary surgery and revision surgery including instrumentation, but without segmenting the vertebrae or adding support in screw placement planning.23 A continuation of that work by Ketcha et al. showed a high accuracy even in patients with severe spinal deformations by using 2D fluoroscopy to map the patient’s anatomy in the OR to a preoperative image set including preoperative plans.7 This solution, like previous similar solutions, effectively moves the planning time out of the OR and decreases the risk of wrong-site surgery as well as extending navigation possibilities where surgical tracking is otherwise not available. However, these systems do not primarily aim to reduce the time spent on navigation or improve workflow by automation, as does the software in our study.

In our experience, the problem of incorrect segmentations and screw suggestions does not translate into patient safety issues because, by intentional design of the system, each suggested screw path needs to be actively confirmed by the planning surgeon. Thus, the screws automatically planned can be viewed as a net positive for reducing navigation time while the misidentified pedicles and screw paths are readily identified and simply require normal screw planning. In our study, the automatic segmentation and screw path suggestion took an average of 11 ± 4 seconds for 5 adjacent vertebrae. In terms of clinical workflow this means a negligible addition to total time, especially compared to time spent planning surgical navigation without automation of screw planning.

Overall, the segmentation system performed well in our study, as evidenced by the results. It is noteworthy that the learning data set included only cadavers without spinal deformities or pathologies. The algorithm proved adaptable enough to compensate for various unexpected pathologies that were not present in the original training set. Similar systems developed for the sole purpose of segmenting and evaluating metastatic disease in vertebrae have previously been proven to overcome the challenge of contour altering spinal disease and be able to still reliably segment the spine,17 but these systems were not developed with the segmentation precision required for the purpose of proposing pedicle screw placement. Further validation of the software used in our study on a set of patients with more diverse spinal pathologies is needed before extending these results to other patient groups. Inclusion of more spinal pathologies in the training sets of the algorithm is the logical next step to enhance clinical precision.

Conclusions

In this study, we have shown that the overall clinically relevant accuracy of a system for automatic vertebral segmentation and pedicle screw planning is 86.1%. By excluding patients with severe spinal deformities and previous surgeries, a clinical accuracy of 95.4% was achieved. This type of technology has the potential to support the surgeon in pedicle screw planning when using surgical navigation.

Disclosures

C.B., J.H., R.N., C.L., D.B., R.H., and M.G. are employed by Philips Research and/or Philips Healthcare, which has a vested interest in the product in this manuscript. The authors report that Karolinska University Hospital and Philips Healthcare have a major collaboration agreement. Dr. Racadio reports that Cincinnati Children’s Hospital Medical Center Department of Radiology has a Master Research Agreement with Philips Medical.

Author Contributions

Conception and design: Burström, Buerger, Nachabe, Babic, Grass, Persson, Edström, Elmi Terander. Acquisition of data: Burström, Buerger, Hoppenbrouwers, Lorenz, Homan, Persson. Analysis and interpretation of data: Burström, Buerger, Hoppenbrouwers, Nachabe, Lorenz, Homan, Persson. Drafting the article: Burström, Buerger, Edström, Elmi Terander. Critically revising the article: Babic, Racadio, Grass, Persson, Edström, Elmi Terander. Reviewed submitted version of manuscript: Burström, Buerger, Hoppenbrouwers, Nachabe, Lorenz, Babic, Racadio, Grass, Persson, Edström, Elmi Terander. Approved the final version of the manuscript on behalf of all authors: Burström. Statistical analysis: Burström, Buerger, Nachabe, Homan. Administrative/technical/material support: Homan. Study supervision: Babic, Racadio, Grass, Edström, Elmi Terander.

Supplemental Information

Previous Presentations

Portions of the data used in this paper have previously been orally presented and included in a non–peer-reviewed conference proceeding (Buerger C, Lorenz C, Babic D, Hoppenbrouwers J, Homan R, Nachabe R, et al: Spine segmentation from C-arm CT data sets: application to region-of-interest volumes for spinal interventions, in SPIE Medical Imaging 2017: Image-Guided Procedures, Robotic Interventions, and Modeling, 2017). The oral presentation took place during a session at the SPIE-sponsored conference SPIE Medical Imaging on February 16, 2017, in Orlando, Florida. Granted this acknowledgment and acknowledgment in the manuscript text, SPIE has given us permission to reuse all material from the proceeding. However, no text or figures are reused herein.

References

  • 1

    Ballard DH: Generalizing the Hough transform to detect arbitrary shapes. Pattern Recognit 13:111122, 1981

  • 2

    Beichel R, Bauer C, Bornik A, Sorantin E, Bischof H: Liver segmentation in CT data: A segmentation refinement approach, in Heimann T, Styner M, van Ginneken B (eds): 3D Segmentation in the Clinic: A Grand Challenge, 2007 (http://mbi.dkfz-heidelberg.de/grand-challenge2007/sites/proceed.htm) [Accessed January 21, 2019]

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 3

    Buerger C, Lorenz C, Babic D, Hoppenbrouwers J, Homan R, Nachabe R, et al.: Spine segmentation from C-arm CT data sets: application to region-of-interest volumes for spinal interventions. Proc SPIE 10135:101351N, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 4

    Byrnes TJ, Barrick TR, Bell BA, Clark CA: Semiautomatic tractography: motor pathway segmentation in patients with intracranial vascular malformations. Clinical article. J Neurosurg 111:132140, 2009

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 5

    Cohen J: A coefficient of agreement for nominal scales. Educ Psychol Meas 20:3746, 1960

  • 6

    De Leener B, Cohen-Adad J, Kadoury S: Automatic segmentation of the spinal cord and spinal canal coupled with vertebral labeling. IEEE Trans Med Imaging 34:17051718, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 7

    De Silva T, Uneri A, Ketcha MD, Reaungamornrat S, Goerres J, Jacobson MW, et al.: Registration of MRI to intraoperative radiographs for target localization in spinal interventions. Phys Med Biol 62:684701, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 8

    Drazin D, Kim TT, Polly DW Jr, Johnson JP: Introduction. Intraoperative spinal imaging and navigation. Neurosurg Focus 36(3):Introduction, 2014

  • 9

    Du JP, Fan Y, Wu QN, Wang DH, Zhang J, Hao DJ: Accuracy of pedicle screw insertion among 3 image-guided navigation systems: systematic review and meta-analysis. World Neurosurg 109:2430, 2018

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 10

    Ecabert O, Peters J, Schramm H, Lorenz C, von Berg J, Walker MJ, et al.: Automatic model-based segmentation of the heart in CT images. IEEE Trans Med Imaging 27:11891201, 2008

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 11

    Elmi Terander A, Burström G, Nachabe R, Skulason H, Pedersen K, Fagerlund M, et al.: Pedicle screw placement with augmented reality surgical navigation with intraoperative 3D imaging: a first in-human prospective cohort study. Spine (Phila Pa 1976) [epub ahead of print], 2018

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 12

    Elmi-Terander A, Skulason H, Söderman M, Racadio J, Homan R, Babic D, et al.: Surgical navigation technology based on augmented reality and integrated 3D intraoperative imaging: a spine cadaveric feasibility and accuracy study. Spine (Phila Pa 1976) 41:E1303E1311, 2016

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 13

    Gelalis ID, Paschos NK, Pakos EE, Politis AN, Arnaoutoglou CM, Karageorgos AC, et al.: Accuracy of pedicle screw placement: a systematic review of prospective in vivo studies comparing free hand, fluoroscopy guidance and navigation techniques. Eur Spine J 21:247255, 2012

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 14

    Goerres J, Uneri A, De Silva T, Ketcha M, Reaungamornrat S, Jacobson M, et al.: Spinal pedicle screw planning using deformable atlas registration. Phys Med Biol 62:28712891, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 15

    Härtl R, Lam KS, Wang J, Korge A, Kandziora F, Audigé L: Worldwide survey on the use of navigation in spine surgery. World Neurosurg 79:162172, 2013

  • 16

    Helm PA, Teichman R, Hartmann SL, Simon D: Spinal navigation and imaging: history, trends, and future. IEEE Trans Med Imaging 34:17381746, 2015

  • 17

    Hojjat SP, Hardisty MR, Whyne CM: Micro-computed tomography-based highly automated 3D segmentation of the rat spine for quantitative analysis of metastatic disease. J Neurosurg Spine 13:367370, 2010

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 18

    Huang J, Jian F, Wu H, Li H: An improved level set method for vertebra CT image segmentation. Biomed Eng Online 12:48, 2013

  • 19

    Kim Y, Kim D: A fully automatic vertebra segmentation method using 3D deformable fences. Comput Med Imaging Graph 33:343352, 2009

  • 20

    Klinder T, Ostermann J, Ehm M, Franz A, Kneser R, Lorenz C: Automated model-based vertebra detection, identification, and segmentation in CT images. Med Image Anal 13:471482, 2009

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 21

    Korez R, Ibragimov B, Likar B, Pernuš F, Vrtovec T: A framework for automated spine and vertebrae interpolation-based detection and model-based segmentation. IEEE Trans Med Imaging 34:16491662, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 22

    Kosmopoulos V, Schizas C: Pedicle screw placement accuracy: a meta-analysis. Spine (Phila Pa 1976) 32:E111E120, 2007

  • 23

    Lo SF, Otake Y, Puvanesarajah V, Wang AS, Uneri A, De Silva T, et al.: Automatic localization of target vertebrae in spine surgery: clinical evaluation of the LevelCheck registration algorithm. Spine (Phila Pa 1976) 40:E476E483, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 24

    Mandell JG, Langelaan JW, Webb AG, Schiff SJ: Volumetric brain analysis in neurosurgery: Part 1. Particle filter segmentation of brain and cerebrospinal fluid growth dynamics from MRI and CT images. J Neurosurg Pediatr 15:113124, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 25

    Mason A, Paulsen R, Babuska JM, Rajpal S, Burneikiene S, Nelson EL, et al.: The accuracy of pedicle screw placement using intraoperative image guidance systems. J Neurosurg Spine 20:196203, 2014

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 26

    Moses ZB, Mayer RR, Strickland BA, Kretzer RM, Wolinsky JP, Gokaslan ZL, et al.: Neuronavigation in minimally invasive spine surgery. Neurosurg Focus 35(2):E12, 2013

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 27

    Otake Y, Schafer S, Stayman JW, Zbijewski W, Kleinszig G, Graumann R, et al.: Automatic localization of vertebral levels in x-ray fluoroscopy using 3D-2D registration: a tool to reduce wrong-site surgery. Phys Med Biol 57:54855508, 2012

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 28

    Overley SC, Cho SK, Mehta AI, Arnold PM: Navigation and robotics in spinal surgery: where are we now? Neurosurgery 80(3S):S86–S99, 2017

  • 29

    Peters J, Ecabert O, Meyer C, Kneser R, Weese J: Optimizing boundary detection via simulated search with applications to multi-modal heart segmentation. Med Image Anal 14:70–84, 2010

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 30

    Rajasekaran S, Vidyadhara S, Ramesh P, Shetty AP: Randomized clinical study to compare the accuracy of navigated and non-navigated thoracic pedicle screws in deformity correction surgeries. Spine (Phila Pa1976) 32:E56E64, 2007

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 31

    Saddi KA, Rousson M, Chefd’hotel C, Cheriet F: Global-to-local shape matching for liver segmentation in CT imaging, in Heimann T, Styner M, van Ginneken B (eds): 3D Segmentation in the Clinic: A Grand Challenge, 2007 (http://mbi.dkfz-heidelberg.de/grand-challenge2007/sites/proceed.htm) [Accessed January 21, 2019]

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 32

    Scarone P, Vincenzo G, Distefano D, Del Grande F, Cianfoni A, Presilla S, et al.: Use of the Airo mobile intraoperative CT system versus the O-arm for transpedicular screw fixation in the thoracic and lumbar spine: a retrospective cohort study of 263 patients. J Neurosurg Spine 29:397406, 2018

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 33

    Shin BJ, James AR, Njoku IU, Härtl R: Pedicle screw navigation: a systematic review and meta-analysis of perforation risk for computer-navigated versus freehand insertion. J Neurosurg Spine 17:113122, 2012

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 34

    Xiao R, Miller JA, Sabharwal NC, Lubelski D, Alentado VJ, Healy AT, et al.: Clinical outcomes following spinal fusion using an intraoperative computed tomographic 3D imaging system. J Neurosurg Spine 26:628637, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 35

    Yang Z, Jin Z: Modeling and specifying parametric adaptation mechanism for self-adaptive systems, in Zowghi D, Jin Z (eds): Requirements Engineering. Berlin: Springer, 2014, pp 105119

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 36

    Zhang W, Takigawa T, Wu Y, Sugimoto Y, Tanaka M, Ozaki T: Accuracy of pedicle screw insertion in posterior scoliosis surgery: a comparison between intraoperative navigation and preoperative navigation techniques. Eur Spine J 26:17561764, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • Collapse
  • Expand
  • FIG. 1.

    A visualization of how the software identifies the vertebrae by a 3-step process, showing the application of the generalized Hough transform (A), the parametric adaptation improving the global alignment (B), and the deformable adaptation finalizing the local alignment (C). Figure is available in color online only.

  • FIG. 2.

    Visualization of the automatic pedicle identification and screw placement suggestion in axial (A) and sagittal (B) views and the visualization of the segmented vertebral section including screw suggestion (C). Figure is available in color online only.

  • FIG. 3.

    Illustration of midpoint measurements. The distance (in red) from the midpoint of each pedicle (in brown) to the midpoint of the corresponding segmented 3D reconstruction of the pedicle (in blue) in the same coronal plane was measured for all segmented pedicles. Figure is available in color online only.

  • FIG. 4.

    The number of segmented pedicles in each patient separated by correct or incorrect identification. Patients with any of the exclusion criteria present are indicated by pound signs (#). Figure is available in color online only.

  • FIG. 5.

    Distribution of distances between the algorithm-suggested pedicle midpoint and the actual manually defined pedicle midpoint. Deviation distances are stratified by correct or incorrect pedicle identification and median (solid vertical line) and quartiles (dashed vertical lines) are highlighted. Figure is available in color online only.

  • 1

    Ballard DH: Generalizing the Hough transform to detect arbitrary shapes. Pattern Recognit 13:111122, 1981

  • 2

    Beichel R, Bauer C, Bornik A, Sorantin E, Bischof H: Liver segmentation in CT data: A segmentation refinement approach, in Heimann T, Styner M, van Ginneken B (eds): 3D Segmentation in the Clinic: A Grand Challenge, 2007 (http://mbi.dkfz-heidelberg.de/grand-challenge2007/sites/proceed.htm) [Accessed January 21, 2019]

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 3

    Buerger C, Lorenz C, Babic D, Hoppenbrouwers J, Homan R, Nachabe R, et al.: Spine segmentation from C-arm CT data sets: application to region-of-interest volumes for spinal interventions. Proc SPIE 10135:101351N, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 4

    Byrnes TJ, Barrick TR, Bell BA, Clark CA: Semiautomatic tractography: motor pathway segmentation in patients with intracranial vascular malformations. Clinical article. J Neurosurg 111:132140, 2009

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 5

    Cohen J: A coefficient of agreement for nominal scales. Educ Psychol Meas 20:3746, 1960

  • 6

    De Leener B, Cohen-Adad J, Kadoury S: Automatic segmentation of the spinal cord and spinal canal coupled with vertebral labeling. IEEE Trans Med Imaging 34:17051718, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 7

    De Silva T, Uneri A, Ketcha MD, Reaungamornrat S, Goerres J, Jacobson MW, et al.: Registration of MRI to intraoperative radiographs for target localization in spinal interventions. Phys Med Biol 62:684701, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 8

    Drazin D, Kim TT, Polly DW Jr, Johnson JP: Introduction. Intraoperative spinal imaging and navigation. Neurosurg Focus 36(3):Introduction, 2014

  • 9

    Du JP, Fan Y, Wu QN, Wang DH, Zhang J, Hao DJ: Accuracy of pedicle screw insertion among 3 image-guided navigation systems: systematic review and meta-analysis. World Neurosurg 109:2430, 2018

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 10

    Ecabert O, Peters J, Schramm H, Lorenz C, von Berg J, Walker MJ, et al.: Automatic model-based segmentation of the heart in CT images. IEEE Trans Med Imaging 27:11891201, 2008

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 11

    Elmi Terander A, Burström G, Nachabe R, Skulason H, Pedersen K, Fagerlund M, et al.: Pedicle screw placement with augmented reality surgical navigation with intraoperative 3D imaging: a first in-human prospective cohort study. Spine (Phila Pa 1976) [epub ahead of print], 2018

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 12

    Elmi-Terander A, Skulason H, Söderman M, Racadio J, Homan R, Babic D, et al.: Surgical navigation technology based on augmented reality and integrated 3D intraoperative imaging: a spine cadaveric feasibility and accuracy study. Spine (Phila Pa 1976) 41:E1303E1311, 2016

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 13

    Gelalis ID, Paschos NK, Pakos EE, Politis AN, Arnaoutoglou CM, Karageorgos AC, et al.: Accuracy of pedicle screw placement: a systematic review of prospective in vivo studies comparing free hand, fluoroscopy guidance and navigation techniques. Eur Spine J 21:247255, 2012

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 14

    Goerres J, Uneri A, De Silva T, Ketcha M, Reaungamornrat S, Jacobson M, et al.: Spinal pedicle screw planning using deformable atlas registration. Phys Med Biol 62:28712891, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 15

    Härtl R, Lam KS, Wang J, Korge A, Kandziora F, Audigé L: Worldwide survey on the use of navigation in spine surgery. World Neurosurg 79:162172, 2013

  • 16

    Helm PA, Teichman R, Hartmann SL, Simon D: Spinal navigation and imaging: history, trends, and future. IEEE Trans Med Imaging 34:17381746, 2015

  • 17

    Hojjat SP, Hardisty MR, Whyne CM: Micro-computed tomography-based highly automated 3D segmentation of the rat spine for quantitative analysis of metastatic disease. J Neurosurg Spine 13:367370, 2010

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 18

    Huang J, Jian F, Wu H, Li H: An improved level set method for vertebra CT image segmentation. Biomed Eng Online 12:48, 2013

  • 19

    Kim Y, Kim D: A fully automatic vertebra segmentation method using 3D deformable fences. Comput Med Imaging Graph 33:343352, 2009

  • 20

    Klinder T, Ostermann J, Ehm M, Franz A, Kneser R, Lorenz C: Automated model-based vertebra detection, identification, and segmentation in CT images. Med Image Anal 13:471482, 2009

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 21

    Korez R, Ibragimov B, Likar B, Pernuš F, Vrtovec T: A framework for automated spine and vertebrae interpolation-based detection and model-based segmentation. IEEE Trans Med Imaging 34:16491662, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 22

    Kosmopoulos V, Schizas C: Pedicle screw placement accuracy: a meta-analysis. Spine (Phila Pa 1976) 32:E111E120, 2007

  • 23

    Lo SF, Otake Y, Puvanesarajah V, Wang AS, Uneri A, De Silva T, et al.: Automatic localization of target vertebrae in spine surgery: clinical evaluation of the LevelCheck registration algorithm. Spine (Phila Pa 1976) 40:E476E483, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 24

    Mandell JG, Langelaan JW, Webb AG, Schiff SJ: Volumetric brain analysis in neurosurgery: Part 1. Particle filter segmentation of brain and cerebrospinal fluid growth dynamics from MRI and CT images. J Neurosurg Pediatr 15:113124, 2015

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 25

    Mason A, Paulsen R, Babuska JM, Rajpal S, Burneikiene S, Nelson EL, et al.: The accuracy of pedicle screw placement using intraoperative image guidance systems. J Neurosurg Spine 20:196203, 2014

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 26

    Moses ZB, Mayer RR, Strickland BA, Kretzer RM, Wolinsky JP, Gokaslan ZL, et al.: Neuronavigation in minimally invasive spine surgery. Neurosurg Focus 35(2):E12, 2013

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 27

    Otake Y, Schafer S, Stayman JW, Zbijewski W, Kleinszig G, Graumann R, et al.: Automatic localization of vertebral levels in x-ray fluoroscopy using 3D-2D registration: a tool to reduce wrong-site surgery. Phys Med Biol 57:54855508, 2012

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 28

    Overley SC, Cho SK, Mehta AI, Arnold PM: Navigation and robotics in spinal surgery: where are we now? Neurosurgery 80(3S):S86–S99, 2017

  • 29

    Peters J, Ecabert O, Meyer C, Kneser R, Weese J: Optimizing boundary detection via simulated search with applications to multi-modal heart segmentation. Med Image Anal 14:70–84, 2010

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 30

    Rajasekaran S, Vidyadhara S, Ramesh P, Shetty AP: Randomized clinical study to compare the accuracy of navigated and non-navigated thoracic pedicle screws in deformity correction surgeries. Spine (Phila Pa1976) 32:E56E64, 2007

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 31

    Saddi KA, Rousson M, Chefd’hotel C, Cheriet F: Global-to-local shape matching for liver segmentation in CT imaging, in Heimann T, Styner M, van Ginneken B (eds): 3D Segmentation in the Clinic: A Grand Challenge, 2007 (http://mbi.dkfz-heidelberg.de/grand-challenge2007/sites/proceed.htm) [Accessed January 21, 2019]

    • PubMed
    • Search Google Scholar
    • Export Citation
  • 32

    Scarone P, Vincenzo G, Distefano D, Del Grande F, Cianfoni A, Presilla S, et al.: Use of the Airo mobile intraoperative CT system versus the O-arm for transpedicular screw fixation in the thoracic and lumbar spine: a retrospective cohort study of 263 patients. J Neurosurg Spine 29:397406, 2018

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 33

    Shin BJ, James AR, Njoku IU, Härtl R: Pedicle screw navigation: a systematic review and meta-analysis of perforation risk for computer-navigated versus freehand insertion. J Neurosurg Spine 17:113122, 2012

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 34

    Xiao R, Miller JA, Sabharwal NC, Lubelski D, Alentado VJ, Healy AT, et al.: Clinical outcomes following spinal fusion using an intraoperative computed tomographic 3D imaging system. J Neurosurg Spine 26:628637, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 35

    Yang Z, Jin Z: Modeling and specifying parametric adaptation mechanism for self-adaptive systems, in Zowghi D, Jin Z (eds): Requirements Engineering. Berlin: Springer, 2014, pp 105119

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation
  • 36

    Zhang W, Takigawa T, Wu Y, Sugimoto Y, Tanaka M, Ozaki T: Accuracy of pedicle screw insertion in posterior scoliosis surgery: a comparison between intraoperative navigation and preoperative navigation techniques. Eur Spine J 26:17561764, 2017

    • Crossref
    • PubMed
    • Search Google Scholar
    • Export Citation

Metrics

All Time Past Year Past 30 Days
Abstract Views 6334 597 0
Full Text Views 608 168 36
PDF Downloads 633 140 21
EPUB Downloads 0 0 0