Utility of virtual monoenergetic images from spectral detector computed tomography in improving image segmentation for purposes of 3D printing and modeling
- 331 Downloads
One of the key steps in generating three-dimensional (3D) printed models in medicine is segmentation of radiologic imaging. The software tools used for segmentation may be automated, semi-automated, or manual which rely on differences in material density, attenuation characteristics, and/or advanced software algorithms. Spectral Detector Computed Tomography (SDCT) is a form of dual energy computed tomography that works at the detector level to generate virtual monoenergetic images (VMI) at different energies/ kilo-electron volts (keV). These VMI have varying contrast and attenuation characteristics relative to material density. The purpose of this pilot project is to explore the use of VMI in segmentation for medical 3D printing in four separate clinical scenarios. Cases were retrospectively selected based on varying complexity, value of spectral data, and across multiple clinical disciplines (Vascular, Cardiology, Oncology, and Orthopedic).
In all four clinical cases presented, the segmentation process was qualitatively reported as easier, faster, and increased the operator’s confidence in obtaining accurate anatomy. All cases demonstrated a significant difference in the calculated Hounsfield Units between conventional and VMI data at the level of targeted segmentation anatomy. Two cases would not have been feasible for segmentation and 3D printing using conventional images only. VMI data significantly reduced conventional CT artifacts in one of the cases.
Utilization of VMI from SDCT can improve and assist the segmentation of target anatomy for medical 3D printing by enhancing material contrast and decreasing CT artifact.
Keywords3D printing Spectral detector CT Segmentation Dual-energy CT Dual layer CT
Institutional Review Board
Philips IntelliSpace Portal
Left Atrial Appendage
Region of Interest
Spectral Detector Computed Tomography
Standard Tessellation Language
Transcatheter Aortic Valve Replacement
Virtual Monoenergetic Images
The workflow for medical three-dimensional (3D) printing is consistent and involves initial acquisition of the imaging data, segmentation of anatomy, 3D mesh post-processing, and physical 3D printing [1, 2]. One of the most time consuming and limiting factors throughout this workflow is at segmentation [3, 4]. Computer software tools such as density thresholding, automatic region growing, edge detection, and manual segmentation rely on a combination of the user’s and software’s ability to differentiate various densities as well as advanced pixel-based software algorithms .
Dual-energy computed tomography has become clinically available in the mid-2000s [6, 7]. There are several different approaches available that can be categorized into emission- and detection-based systems [8, 9]. While the first group comprises several concepts (e.g. dual source computed tomography, kVp-switching computed tomography), there is only one technology clinically available that works on the detector level and does not require protocol selection prior to the scan: Spectral Detector Computed Tomography (SDCT). It is equipped with a dual-layer detector that registers high and low energy photons in the lower and upper, detector layer, respectively .
From a physical perspective, attenuation in CT imaging can be considered as the sum of the photoelectric-effect and Compton-scattering. While the former is predominant in energies up to 100 keV and mostly dependent on the atomic number of any given material, the later depends on the physical density and becomes predominant in energies > 100 keV .Besides other image reconstructions, dual energy image registration allows for computation of so-called virtual monoenergetic images (VMI). These are available in a range of range of 40–200 keV using linear blending and extrapolation of information from both detector layers.
VMI approximate images are acquired with a true monoenergetic X-ray . Hence, low keV VMI accentuate differences regarding the material atomic number resulting in an increase in soft-tissue contrast. This is why iodinated contrast media is frequently administered in CT imaging due to its high atomic number (Z = 53) . In addition, k-edge effects further enhance the iodine-associated attenuation. Photons with an energy in the proximity of an elements k-edge are more likely to be absorbed. Since the k-edge of iodine lies at 33 keV, photons with 40 keV are more likely to be absorbed resulting in increased attenuation . High keV VMI can be used to reduce image artifacts due to photon starvation [13, 14, 15]. These characteristics are highlighted in Additional file 1: Figure S1.
Based on these physics and the dual-energy CT technology available, the combination of conventional and VMI data may facilitate segmentation for 3D printing. This study aimed to investigate if VMI from SDCT accelerates and/or improves image pre-processing and segmentation for medical 3D printing.
Multiple cases were acquired using a clinical SDCT scanner (IQon, Philips Healthcare, Best, The Netherlands) with different imaging protocols. Institutional Review Board (IRB) approval was obtained (NHR-17-57) for retrospective evaluation and four SDCT cases were reviewed for relevant anatomy. Cases were selected based on complexity, value of spectral data compared to associated conventional images, and clinical specialty (Vascular, Cardiology, Oncology, and Orthopedic). Virtual monoenergetic images (VMI) from the original scans at multiple energy levels were obtained. The spectral reconstructions are available for every scan done on the SDCT scanner and all examinations were performed for clinical indications. No scan was acquired for the sole purpose of this study.
Image segmentation was carried out using built-in software tools that come with the vendor’s proprietary image viewer (IntelliSpace Portal (ISP), v9.0, Best, The Netherlands). Both conventional and VMI at 40 to 170 keV at 10 to 20 keV increments were processed and reviewed for segmentation (40 keV, 50 keV, 70 keV, 90 keV, 100 keV, 120 keV, 130 keV, 150 keV, and 170 keV). For cases 1, 2, and 3, the 40 keV images were ultimately utilized due to the closeness to the k-edge value of iodine (33.2 keV) which provided for maximum contrast effect. For case 4, 120 keV data was utilized for metal artifact reduction.
The institution’s single 3D lab senior person completed the requested anatomical segmentation for each of the cases (Additional file 1: Figure S2). Subjective qualitative feedback regarding the performance of segmentation tools was collected by the single person performing the segmentation and differential in calculated Hounsfield Units for relevant anatomy was compared between conventional and VMI.
Additional generation of the Standard Tessellation Language (STL) models was completed in ISP. Postprocessing of the STL files was completed in Autodesk Meshmixer prior to printing. Considerations were made for model orientation and support materials required for 3D printing including removing free-floating elements, proper hollow construction, and maximizing build space. All 3D printed models were made using the Formlabs Form 2 stereolithography (SLA) printer with standard resin material.
Case 1: Transcatheter aortic valve replacement pre-procedure planning for vascular access simulation
Case 2: Left atrial appendage Thrombus segmentation for purposes of patient education
77-year-old male with a history of atrial fibrillation who presented for preprocedural planning for left atrial appendage closure device. Part of the preprocedural workup includes helical cardiac CT with retrospective gating. The scan delay is a threshold/trigger-based hybrid with the scanner manually started once the user sees the first blush of contrast in the right atrium. The patient received 25 mL of iodinated contrast Isovue 370 at 4 mL/sec followed by a saline bolus.
A 3D printed model was requested for patient education. However, the scan had a very washed out appearance and segmentation of the thrombus using conventional CT data only did not yield an accurate model due to limited attenuation differentiation. The corresponding 40 keV VMI from the delayed phase (Fig. 3b) elegantly demonstrates a thrombus in the left atrial appendage. There was a two-fold increase in HU values between the conventional and VMI target anatomy. Using the 40 keV VMI data set, the thrombus volume was well defined, and the resulting 3D printed model was anatomically 1:1 scale (Fig. 3d).
Case 3: Malignant tracheal lesion identification and segmentation for resident physician education
A 3D model was requested for operative planning and resident physician education. Segmentation of the airway with the proximal tracheal lesion was technically feasible on both conventional and VMI data. However, there was a subjective decrease in segmentation time when using VMI. Additionally, there was a subjective increase in confidence when selecting the appropriate margins of the tracheal lesion with respect to the surrounding anatomy due to the significant increase in HU differentiation between conventional and VMI.
Case 4: Salvaged segmentation of left upper extremity trauma with orthopedic hardware artifact
45-year-old-male with history of trauma to the left upper extremity after involvement in a motor vehicle accident. The patient underwent open reduction internal fixation of the left proximal surgical neck and distal humeral shaft fractures with intramedullary rod nailing. After returning ten weeks post-operatively, the patient’s range of motion had improved. However, there was residual left shoulder pain and migration of the proximal hardware screw. A SDCT without contrast of the left upper extremity was performed and demonstrated improved fracture visualization but minimal bony bridging or callus formation.
Hounsfield Unit (HU) values of segmented anatomy for conventional and VMI data from all four presented cases including ratio differential. Comparative and differential calculations were also made between the segmented ROI HU and the adjacent tissue material HU
Conventional HU at ROI (C)
VMI HU at ROI (V)
HU value of tissue surrounding ROI (S)
HU Ratio Difference (C:V)
Δ HU Conventional HU to HU surrounding ROI (C-S)
Δ HU VMI HU to HU surrounding ROI (V-S)
Accurate segmentation of desired anatomy from imaging data for the purposes of 3D printing requires using a combination and variety of software tools [1, 2, 4]. All three of the most common semi-automated segmentation tools, global thresholding, edge detection, and region growing, rely on voxel Hounsfield Unit values to differentiate between different tissues and surrounding anatomical structures . Even manual segmentation tools such as cropping and sculpting count on visual acuity to see the greyscale differences between the target anatomy. SDCT provides both conventional and VMI data with a significant difference in Hounsfield Unit values which aids in the automatic and manual segmentation workflow. While other advanced pixel-based software algorithms that do not completely rely on tissue density may be helpful in certain case scenarios, these tools may not be universally applicable to all cases. To the best of our knowledge, there has been no previously published work associating spectral detector computed tomography and medical 3D printing applications.
In all the cases described above, the segmentation process using VMI data was qualitatively reported as easier, faster, and increased the operator’s confidence in obtaining accurate anatomy. Case 4 also demonstrated how imaging artifacts could be reduced on SDCT enabling segmentation and 3D printing from otherwise unsalvageable imaging sources [13, 14, 15]. The single segmentation operator commented that without the VMI data in each of the four detailed cases, the requested segmentation may not have been possible and would have resulted in either time-prohibiting manual segmentation or rescanning the desired patient anatomy.
Since the SDCT VMI data is automatically acquired with every scan without requiring an additional protocol selection or intervention, the VMI data is available to aid in segmentation any time after the image acquisition. This key factor is what enabled our team to retrospectively review any case completed on the SDCT for VMI segmentation. The only workflow change is prospectively requesting any clinical 3D printing cases to have their imaging acquired on the SDCT which we have done in our practice setting.
Due to the complexities of the SDCT and VMI data along with the limited institutional resources, we were able to recruit only one senior person for segmentation of the cases with the knowledgebase and software skills required. One of the primary goals of this project was to establish the concept and workflow for integrating dual-energy CT VMI data into medical 3D printing segmentation. Future directions for this project include training and recruiting additional persons for segmentation to formally quantify and evaluate variability in the segmentation process between multiple persons.
One of the limitations of this workflow is that the VMI data is optimized for use in Philips Intellispace Portal 9.0. The 3D segmentation tools in Philips ISP are robust and the ability to use a slider/toggle to dynamically switch between different keV settings is extremely helpful for both automatic and manual segmentation. However, if the workflow requires utilization of different or more advanced segmentation software, the VMI DICOM data must be post-processed and exported individually at each desired keV value (40 keV DICOM dataset, 60 keV DICOM dataset, etc.). Future software advancements to integrate VMI data natively in other platforms would be ideal and allow for greater flexibility in utilizing spectral data on multiple platforms.
Including SDCT at the image acquisition phase allows for better utilization of segmentation tools during the medical 3D printing workflow. Complex and previously difficult cases where densities could not be separated are better demarcated with the VMI data. With the rapidly rising adoption of both medical 3D printing and dual-energy computed tomography, combining these technologies may lead to more advanced clinical applications such as new algorithms/tools for segmentation or automatic segmentation of 3D printable regions of interest based on dual-energy material properties. Incorporating SDCT acquisition improves the downstream 3D segmentation process and further enhances the medical 3D printing workflow.
The article processing charge for this publication has been funded generously by an unrestricted educational grant from Stratasys.
The materials used for 3D printing were supported by funding from an internal University Hospitals Cleveland Medical Center Graduate Medical Education Innovation Award.
Availability of data and materials
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
EK, NGH, and AG selected the cases. LC, EK, and AG completed the segmentation, STL mesh postprocessing, and validation of the cases respectively. EK managed the 3D printing and comparative results. EK and NGH prepared the draft manuscript. EK and AG designed the figures. CK, NR, and AG each revised the manuscript. All authors read and approved the final manuscript.
Ethics approval and consent to participate
IRB approval was obtained for utilization of imaging data (UH IRB NHR-17-57).
Consent for publication
None of the authors has any relationship with Stratasys. NGH is on the speaker’s bureau of Philips Healthcare.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
- 13.Große hokamp N, Hellerbach A, Gierich A, et al. Reduction of artifacts caused by deep brain stimulating electrodes in cranial computed tomography imaging by means of virtual Monoenergetic images, metal artifact reduction algorithms, and their combination. Investig Radiol. 2018;53(7):424–31.Google Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.