Automatic Segmentation of Bone Selective MR Images for Visualization and Craniometry of the Cranial Vault

STRUCTURED ABSTRACT

Rationale and Objectives:

Solid-state MRI has been shown to provide a radiation-free alternative imaging strategy to CT. However, manual image segmentation to produce bone-selective MR-based 3D renderings is time and labor intensive, thereby acting as a bottleneck in clinical practice. The objective of this study was to evaluate an automatic multi-atlas segmentation pipeline for use on cranial vault images entirely circumventing prior manual intervention and to assess concordance of craniometric measurements between pipeline produced MRI and CT-based 3D skull renderings.

Materials and Methods:

Dual-RF, dual-echo, 3D UTE pulse sequence MR data were obtained at 3T on 30 healthy subjects along with low-dose CT images between December 2018 to January 2020 for this prospective study. The four-point MRI datasets (two RF pulse widths and two echo times) were combined to produce bone-specific images. CT images were thresholded and manually corrected to segment the cranial vault. CT images were then rigidly registered to MRI using mutual information. The corresponding cranial vault segmentations were then transformed to MRI. The “ground truth” segmentations served as reference for the MR images. Subsequently, an automated multi-atlas pipeline was used to segment the bone-selective images. To compare manually and automatically segmented MR images, the Dice similarity coefficient (DSC) and Hausdorff distance (HD) were computed, and craniometric measurements between CT and automated-pipeline MRI-based segmentations was examined via Lin’s concordance coefficient (LCC).

Results:

Automated segmentation reduced the need for an expert to obtain segmentation. Average DSC was 90.86±1.94%, and average 95^th percentile HD was 1.65±0.44 mm between ground truth and automated segmentations. MR-based measurements differed from CT-based measurements by 0.73–1.2 mm on key craniometric measurements. LCCfor distances between CT and MR-based landmarks were vertex-basion: 0.906, left-right frontozygomatic suture: 0.780, and glabella-opisthocranium: 0.956 for the three measurements.

Conclusion:

Good agreement between CT and automated MR-based 3D cranial vault renderings has been achieved, thereby eliminating the laborious manual segmentation process. Target applications comprise craniofacial surgery as well as imaging of traumatic injuries and masses involving both bone and soft tissue.

INTRODUCTION

In the field of craniofacial surgery, CT imaging is the gold-standard imaging modality for diagnostic and surgical planning purposes. High-resolution 3D CT renderings are used with virtual surgical technology to plan osteotomies, to move anatomic regions with six degrees of freedom, and to evaluate different surgical options (1). However, CT involves radiation, which is known to increase a person’s risk of cancer and cancer-related mortality later in life (2). Children, the main population treated by craniofacial surgeons, are especially susceptible to radiation due to their higher life-time attributable risk (LAR) of malignancy and increased susceptibility to radiation-induced carcinogenesis (3). Preoperative imaging requires high spatial resolution, which steeply increases radiation dosage as well. The field of craniofacial surgery needs an alternative imaging modality that is non-radiative, accurate, rapid, and useful for both diagnostic and surgical planning purposes at all patient ages.

Bone-selective MRI allows for selective acquisition of data from calcified tissues as previously described in an article in this Journal by some of the present authors (4). That paper sought to evaluate agreement between CT and manually segmented MRI, reporting a concordance correlation coefficient of 0.998–1.00 between the two imaging modalities for standard craniometric measurements of the skull. However, a limitation of bone-selective MRI preventing it from supplanting CT imaging in clinical practice is the time and labor-intensive process required to produce manual 3D segmentations. Currently, manual skull segmentation takes about 1.5 hours of time per 3D MR image set by an expert operator with an extensive understanding of cranial vault anatomy.

Multi-atlas segmentation methods (5) have previously been used to successfully segment the hippocampus (6) and the whole brain in MRI (7), and dentition in CT (8). These methods are robust to diverse anatomy across the subject population, and also greatly reduce segmentation time. We hypothesized that utilization of a multi-atlas segmentation pipeline for bone-selective MRI would maintain high concordance between bone-selective MRI and CT while substantially reducing segmentation time and operator effort.

The objective of the present study was to implement and evaluate an automatic multi-atlas segmentation pipeline for use on cranial vault images entirely circumventing prior manual intervention and to assess concordance of craniometric measurements between pipeline produced MRI and CT-based 3D skull renderings.

MATERIALS AND METHODS

The protocol of this prospective study was approved by the authors’ Institutional Review Board (IRB#). All participants provided written informed consent, in compliance with the Health Insurance Portability and Accountability Act.

Study Participants

Data was obtained at The Children’s Hospital of Philadelphia (CHOP). Participants were recruited through a study description emailed to students of The University of Pennsylvania institution who were consecutively enrolled from December 2018 to January 2020 residing on the university campus. Eligible subjects were adults over 18 years of age. Subjects were excluded for ferromagnetic materials anywhere in or on their body (metallic braces, non-removable piercings, prosthetic joints, pacemaker), as well as craniofacial anomalies or prior craniofacial surgeries. Imaging data from 25 new study participants as well as the raw CT and bone-selective MR imaging data of five study participants enrolled in a previous feasibility project were included in this study. In total, 30 healthy adults were enrolled: (n=15 male, n=15 female), race/ethnicity: Caucasian (36.7 %, n=11), African American (10.0%, n=3), Asian (33.3%, n=10), Hispanic (16.7%, n=5), Other (3.3%, n=1), Age: median 26.4 yr (range 21.5–45.0 yr) (Table 1).

TABLE 1.

Demographics of the current study.

Subjects Characteristics	Count (% of the total cohort)
Number of Participants	30
Sex
Male	15 (50.0)
Female	15 (50.0)
Race/Ethnicity
Caucasian	11 (36.7)
Black/African American	3 (10.0)
Asian	10 (33.3)
Hispanic	5 (16.7)
Other	1 (3.3)
Age (median, years)	26.4
Age Range (years)	21.5 – 45.0

Study Protocol

All subjects underwent bone-selective MRI and a reduced-dose CT scan following the pediatric craniofacial imaging protocol. No contrast was used for CT or MR images. Image post-processing, multi-atlas segmentation, automated pipeline accuracy assessment and craniometrics concordance analysis was completed by the first authors (FA1, FA2).

Bone-Selective MRI

Subjects were imaged with a dual-radiofrequency (RF), dual-echo, 3D Ultrashort Echo Time (UTE) pulse sequence using a 3T (Prisma; Siemens Medical Solutions, Erlangen, Germany) scanner at 1.1 mm isotropic voxels. In all MR imaging, a 20-channel head/neck coil was used for signal reception, while prior to actual data acquisitions a short scout scan (~10 seconds) was performed for adjusting image positions.

The 3D UTE pulse sequence used in this study employs a dual-RF and dual-echo configuration to exploit the sensitivity of bone signals to RF pulse duration as well as echo time (TE), thereby allowing for enhanced bone contrast with a proper signal combination (9, 10). Specifically, two hard RF pulses (RF₁, RF₂), which differ in duration and amplitude but impart the same nominal flip angles, are alternately applied with an equidistant time-of-repetition (TR), while within each TR two half-echoes are acquired, one at a short TE (TE₁) and one at a long TE (TE₂), with ramp sampling. In the image reconstruction process, the four sets of half-echoes (pertinent to RF₁TE₁, RF₁TE₂, RF₂TE₁, and RF₂TE₂) are combined via a view-sharing strategy (11), leading to two sets of volumetric 3D head images, I₁ (RF₁TE₁ and RF₂TE₁ data combined) and I₂ (RF₁TE₂ and RF₂TE₂ data combined). Finally, bone-specific images (I_bone) are derived by computing (I₁-I₂)/ (I₁+I₂). See recent publications for a more detailed acquisition protocol (4, 12) and Figure 1 for a qualitative depiction of the three images.

Figure 1.

Reconstructed images: The four sets of half-echoes (pertinent to RF₁TE₁, RF₁TE₂, RF₂TE₁, and RF₂TE₂) are combined resulting in two sets of volumetric 3D head images, A) I₁; B) I_{2 :} C) bone-specific (I_bone). The latter images are computed as (I₁-I₂)/ (I₁+I₂). For additional detail on the acquisition protocol, see (4,9,11).

MRI scan parameters:

TR = 7 ms, TE₁/TE₂ = 0.06/2.36 ms, RF1/RF2 durations = 0.04/0.52 ms, nominal flip angle = 12°, readout bandwidth = ±125 kHz, matrix size 256 × 256 × 256, and field of view = 280 × 280 × 280 mm³, number of radial spokes = 25,000, and scan time = 6 minutes. We have applied a bias correction method “N4ITK: Improved N3 Bias Correction” (13) on the MR images.

Reduced-Dose Craniofacial CT Technique and Parameters

CHOP’s standard clinical craniofacial imaging protocol was used with a reduced radiation dose at (0.47–0.53) x (0.47–0.53) mm² in-plane and 0.75 mm slice thickness were acquired. The absorbed radiation dose per CT was in the range of 4–7 mGy with parameters set to not exceed 7mGy, which compares favorably with the typical absorbed adult head CT radiation dosage of 55–70 mGy. A single scanner (Siemens Somatom Definition, Erlangen, Germany) was used for all CT.

Image Post-Processing: CT Segmentation

The MR images were normalized and rescaled between the grayscale values of 0–1000 after setting the lower and upper 0.1 percentile as zero. Each CT image was thresholded between 100–400 Hounsfiled units (HU) and the resulting segmentations were then manually edited by FA1 to only include the cranial vault using the open-source tool ITK-SNAP (6). Facial structures (orbit, maxilla, mandible) and cervical spine below the first vertebra were manually removed. The images were then independently checked by FA2 to remove any spurious voxels. The CT images were rigidly registered to the image I₂ (see Section “Bone-Selective MRI”) using mutual information metric via the open-source registration software „greedy’ (6) and the corresponding cranial vault segmentations were warped to I₂ image space. In this manner, “ground truth”, expert-labeled, segmentations of the cranial vault from CT were obtained as reference to the MR images.

Multi-Atlas Automated Segmentation

An automated multi-atlas segmentation pipeline (Figure 2) (11) was used to segment the cranial vault in the I_bone images. The pipeline involves two steps: training and segmentation. The training step yields a dataset called an „atlas package’. Each pair of registered I₂ and I_bone images, along with the corresponding ground truth segmentations, forming a set of atlases, serves as input to the training step. A series of operations are performed: groupwise deformable registration of all I₂ images generates an unbiased population template (14) and pairwise registration between all images in template space. The unbiased population template is constructed using the Advanced Normalization Tools (ANTs) affine and high-dimensional deformable registration algorithms. These algorithms are based on symmetric diffeomorphic image registration with cross-correlation (15) and the diffeomorphic image averaging approach (16, 17) which alternates between averaging the intensity of all images and registering all images to the intensity average. Multi-atlas segmentation is performed in a leave-one-out manner on the set of atlases.

Figure 2.

Graphical illustration of the training and segmentation stages in the automated segmentation pipeline. The training pipeline takes as its input a set of “atlas” datasets, each consisting of a I₂ and I_bone MRI scans of the same subject, and a ground truth segmentation. The training pipeline outputs an “atlas package,” which is then used as the input to the segmentation pipeline. The segmentation pipeline uses the atlas package to automatically label the I₂ MRI of a new subject, using that subject’s I_bone as an additional input. See Yushkevich et al. (12) for details.

The segmentation step automatically segments the cranial vault of new subjects. The inputs to the segmentation step are the I₂ and I_bone images of the new subject, and the atlas package created in the training step. The segmentation procedure is as follows: The I₂ image of the new subject is registered to the unbiased population template contained in the atlas package. The deformation fields obtained by this registration are used to resample the I₂ and I_bone images into the template space. Within each template, each of the I_bone images in the atlas package is registered to the I_bone image of the new subject. The ground truth segmentations of the I_bone images in the atlas package are mapped into the space of the new subject’s I_bone image. A consensus multi-atlas segmentation of the new subject’s I_bone image is computed using joint label fusion (JLF) (18) and refined by AdaBoost classifier (18).

Craniometric Measurements

Segmentations were rendered in three dimensions using Paraview software (19). Six craniometrics landmarks (glabella, opisthocranion, right and left frontozygomatic suture, vertex, and basion) were identified on each adult cranial vault rendering and three craniometrics distance measurements (glabella to opisthocranion, left frontozygomatic suture to right frontozygomatic suture, and vertex to basion) were taken using the Paraview ruler tool (Table 2). Each distance was measured by a single assessor (FA1). See Figure 7 for a qualitative depiction of the three measurements. Lin’s concordance correlation (LCC) test was applied to assess agreement between mean measurements obtained from MR-based and CT-based 3D skull renderings on STATA 15.1 (StataCorp, LP, College Station, Texas) (20, 21).

TABLE 2.

Craniometric Lin’s concordance correlation scores between segmentations obtained by thresholded and then manually corrected CT segmentations and automated MRI segmentations. SD: standard deviation.

Measurement	CT Distance (mm ± SD)	Automated MRI Distance (mm ± SD)	Mean Difference (mm ± SD)	Mean percent Difference (% ± SD)	Lin’s Concordance Correlation	95% Confidence Interval
Vertex-Basion	140.3 ± 6.0	141.5 ± 5.1	1.2 ±2.1	0.86 ± 1.50	0.906	0.846–0.966
Left-right frontozygomatic sutures	102.6 ± 5.8	103.8 ± 5.0	1.2 ± 3.4	1.17 ± 3.31	0.780	0.642–0.918
Glabella-Opisthocranion	182.1 ± 9.6	181.4± 10.3	0.73 ± 2.9	0.4 ± 1.59	0.956	0.925–0.987

Figure 7.

Three craniometric measurements formed from six craniometric landmarks as indicated on CT 3D renderings and shown from the data of one subject:

1. Glabella-Opisthocranion (G-O)

   • Glabella: The most prominent point on the mid-sagittal plane between supraorbital ridges
   • Opisthocranion: The most prominent posterior point on the occiput
2. Left and right frontozygomatic sutures (FZ-FZ)

   • Frontozygomatic suture: The most lateral point on the parietal bone orthogonal to sagittal plane
3. Vertex-Basion (V-B)

   • Vertex: Highest point on skull in sagittal plane with head in the Frankfort horizontal position
   • Basion: Midpoint of the anterior margin of the foramen magnum, at the junction of the outer and inner tables of the vault.

RESULTS

Assessment of Automated Segmentation

The automated segmentation for each subject was compared with the ground truth expert-labeled segmentations, as described in Section “Post-Image Processing: Manual CT Segmentation”, using two measurements: Dice similarity coefficient (22), a volumetric relative overlap measure, and 95^th percentile Hausdorff distance (HD) (23), which computes the distance between surfaces of the automatic and ground truth segmentations.

Accuracy of Automated Segmentation

We evaluated the method using leave-one-out cross validation setting. The average ± standard deviation DSC across all the subjects was 90.86 ± 1.94 %, and the 95^th percentile HD was 1.65 ± 0.44 mm. Figures 3–5 show qualitative automated segmentation 3D renderings and 2D slices compared with the expert-labeled ground truth segmentations. In Figure 3B and 3C, the ground truth and the automated segmentations are superimposed on the MR images. Figure 4 further shows a stack of slices through the three viewing planes with the corresponding automated segmentations overlaid. The superimposition shows a clear delineation of the cranial vault in the automated segmentations and resembles closely when visually compared with the ground truth segmentations. Figure 5C shows 3D renderings of the automatically segmented cranial vault superimposed on the ground truth segmentation for a sample subject. The most notable differences between automatic segmentations and ground truth segmentations seem to be at the lateral orbital rim and cranial vertex. Figure 6 shows the Bland-Altman plots which depicts that the measures do not differ significantly from the mean difference for the two measurements of automated and ground truth segmentations, and thereby imply no bias is present.

Figure 3.

2D slice views in a typical study subject. Columns A: Image I₂; B: ground truth segmentations superimposed on I₂ images is in red; C: segmented bone superimposed on I₂ images is in red.

Figure 5.

3D renderings comparing automated segmentation with the ground truth (CT) for one of the subjects. Shown are three viewing planes; lateral, posterior, and anterior for A) ground truth segmentation in blue, B) automated segmentation in yellow, and C) automated segmentation overlaid onto ground truth segmentation.

Figure 4.

Automatically segmented bone images from 3D stack superimposed on the I₂ image for one of the study subjects.

Figure 6.

Bland-Altman plots for the three distance measurements (in mm) illustrated in Fig.6. The difference between the automated and the ground truth segmentation is shown on the y-axis (vertical), the average of automated and ground truth segmentation is given on x-axis (horizontal). A) Glabella-Opisthocranion, B) left and right frontozygomatic, C) Vertex-Basion. The mean difference is the estimated bias, and the standard deviation of the differences measures the random fluctuations around this mean. The measures do not differ significantly from the mean difference, and hence no bias is present.

Assessment of Craniometric Measurements Accuracy

The mean distance across all subject CT segmentations were vertex-basion (V-B): 140.3 ± 6.0 mm, left frontozygomatic suture left - right frontozygomatic suture (left FZ-right FZ): 102.6 ± 5.8 mm, and glabella-opisthocranium (G-O): 182.1 ± 9.6 mm, and for the automated MR segmentations V-B: 141.5 ± 5.1 mm, L_FZ-R_FZ: 103.8 ± 5.0 mm, and G-O: 181.4 ± 10.3 mm. Thus, MR-based and CT-based measurements differed by mean distances differences ranging from 0.73mm – 1.2 mm and mean percent differences ranging from 0.4–1.17%. Lin’s concordance correlation coefficient for the distances between CT and MR-based landmarks were 0.906 (95% Confidence Interval (CI): 0.846–0.966), 0.780 (95% CI: 0.642–0.918), and 0.956 (95% CI: 0.925–0.987) for the three measurements across the cohort (Table 2). Figure 6 shows the Bland-Altman plots for the three described craniometric distances.

DISCUSSION

The present study demonstrates that good agreement is achievable between the gold standard (thin-slice CT) and bone-selective MRI through use of a multi-atlas automated segmentation pipeline, thereby eliminating the need for an expert to obtain measurements. A previously published feasibility study on this imaging methodology showed high concordance but clinical applicability was limited by long manual segmentation times of 1–1.5 hours per case (4). The described segmentation pipeline obviates the need for an expert operator, in turn.

Non-radiative alternatives to CT are of particular interest to craniofacial surgeons due to the numerous preoperative and postoperative CT scans their patients receive. An examination of the diagnosis and treatment of patients with craniosynostosis provides insight into the high CT utilization in this field. When an infant present with an abnormal head shape, potentially indicative of a premature suture fusion, low-dose CT with 3D reconstruction is ordered to confirm the diagnosis, assess patency of additional sutures, evaluate for intracranial pathology, and to use in planning the corrective operation (24, 25). Additional CT images are acquired in the immediate postoperative period and as needed later in childhood to assess additional operative requirements. Further, CT scans may be ordered for patients with syndromic craniosynostosis to plan for and assess the outcome of midface and orbital reconstruction. CT imaging requirements for treatment of hemifacial microsomnia, Treacher-Collins Syndrome, and Pierre Robin Sequence are similarly demanding. Due to the risks of ionizing radiation, concerns have been raised by craniofacial surgeons dissuading the use of CT scans in clear cases of single suture craniosynostosis (25). Though CT reduction guidelines have been suggested (26) and lower radiation imaging protocols have been implemented (27), no high-quality alternatives to CT imaging, such as those demonstrated in the present work, have been adopted. Cranial ultrasound is a sensitive option for the diagnosis of craniosynostosis in infants but is inconclusive after one year of age (28).

Non-radiative imaging alternatives are not only necessary in the field of craniofacial surgery, they should be sought throughout much of pediatric and adult medicine. The ability to simultaneously capture signals from soft and hard tissues (achievable with the new MRI technology) would be notably relevant in cases of head trauma as well as extremity reconstruction, musculoskeletal trauma, and visualization of masses involving bone and soft tissue. In the fields of pediatric orthopedic surgery (29), neurosurgery (30), and pediatric thoracic and abdominal radiology (31) scientists are seeking alternatives to standard CT imaging techniques. Bone-selective MRI with multi-atlas segmentation could be applicable in these situations as well. The multi atlas-based segmentation pipeline could be applied directly to a new dataset by obtaining some manually labelled segmentations. To segment the pediatric population, we would segment images to obtain „ground truth’ segmentations to get the training data. The trained pipeline could then be tested on the held-out dataset as described in the study

Craniometric measurement comparisons between MR pipeline and CT based segmentations showed substantial variation based on the specific measurement performed. The G-O measurement showed a high degree of correlation (0.956), V-B measurements were less strongly correlated (0.906) and the FZ-FZ measurement yielded relatively poor correlation of 0.780. However, even for FZ-FZ, the measurement with the lowest degree of correlation, the mean difference in distance was small (1.2 ± 3.4 mm). Virtual surgical planning, a technique commonly utilized for craniofacial and orthognathic surgery, has been reported by some groups to have greater error than observed here (1.98 mm linear error, 1.19° angular error) (32). Multiple studies comparing craniofacial virtual planning to on-table results have determined errors under 2mm as the threshold for clinical significance (33). The 95^th percentile Hausdorff distance is well within this 2mm threshold. Of note, orthognathic surgery is an area in which great precision is required to correct jaw occlusion and cant. Other surgeries likely would allow a greater degree of error than 2 mm while maintaining surgical accuracy. For the other two measurements (V-B, G-O) the mean percent difference in distance between MR and CT was less than 1%. These results suggest notably small measured and percent differences between cranial vault measurements from CT relative to automated MRI.

The vertex, basion, glabella, and oposthocranion are specific craniometric points while the frontozygomatic suture is not a specific point, but rather a line between the frontal and zygomatic bones. An attempt was made to select the most central portion of the suture in the sagittal plane. However, the lack of a specific easily identifiable point introduces user bias, which may have falsely reduced the correlation of our results. Though the frontozygomatic suture was identifiable on MRI for all patients, it was more difficult to visualize on automated pipeline MRI than CT. This is consistent with findings from Zhang et al. on five adult subjects (4). As the authors noted, this discrepancy in suture visualization is likely due to the lower in-plane resolution of MRI (1.1 mm) as compared to CT (0.47 – 0.53 mm). This has not limited adequate suture visualization on bone-selective MRI in other studies and is unlikely to be a limitation of the diagnostic potential of bone-selective MRI (24). The frontozygomatic suture is also closer to air-tissue interfaces than the other areas measured. The dual-echo-based MRI method applied in this study is limited by local magnetic susceptibility differences causing small errors due to the resulting induced frequency shifts, such as that seen between air and tissue water (11) because of the combination in processing of images involving TE~2ms with those at TE = 60µs (where phase errors at magnetic susceptibility interfaces are negligible). Additionally, as compared to other studied bone regions, the frontozygomatic bone is thinner and has been associated with partial volume errors (4).

Though bone-selective MRI has previously shown to be accurate, the manual segmentation time required was a limitation for acceptability in clinical practice. For certain surgical problems, such as the treatment of facial trauma, rapid imaging and surgical intervention is especially important. In the case of orbital blowout fractures, for instance, CT imaging is ordered in the emergency department to examine bony structures (34). If a “trapdoor fracture” (inferior rectus muscle compressed within fracture) is identified, surgical repair is performed within 48 hours of injury. On the other hand, 3D CT has failed to demonstrate soft tissue orbital injuries (fat and muscle entrapment, hematomas) (3). Bone-selective MRI with automated segmentation could be used in this instance to rapidly diagnose both bony and soft-tissue injuries in a single imaging session and 3D segmentations could feasibly be used for surgical planning and producing custom implants within the required surgical timeline (35). The reduced segmentation time through our automated segmentation pipeline would allow bone-selective MRI to be used in clinical practice without causing a delay in treatment even in the case of emergency surgery.

Some limitations of the current segmentation pipeline are noted. The manually delineated data forms the basis of atlas-based segmentation. The number and quality of ground truth cases can critically impact segmentation accuracy. Therefore, standardization of annotation guidelines is vital to obtain accurate automatic segmentations, especially when the atlases are manually segmented by multiple experts. The current ground truth segmentations are obtained by thresholding the CT images, followed by warping the segmentations to MR image space. Therefore, the current study is limited in its comparison of manual segmentation obtained directly from MR images with automated segmentation. Additionally, extension of the method to segment the anterior aspect of the skull, including facial bone, would substantially enhance the method’s potential. The orbit and mandible were not analyzed at this time due to a primary interest in cranial suture visualization and concern that the very thin bone in this area could lead to partial volume effects. We anticipate performing additional analyses to assess the efficacy of bone selective MRI with automated segmentation pipeline for imaging and segmentation of delicate facial structures including the orbits, mandible, nasal bone, vomer, and zygoma in the future. As inter-reader agreement was not the objective of this work, craniometric analysis was performed by one trained reader only under the mentorship of a senior craniofacial surgeon. The authors concede this as a possible limitation.

Certain limitations pertain to image acquisition due to voxel size, which will likely remain larger than CT as spatial resolution largely determines the signal-to-noise ratio (SNR) in MRI. A possible solution is to use a post-processing method, “subvoxel processing” (36), for increasing the apparent resolution of MRI. While SNR can be augmented by increasing total sampling time, this would adversely impact total examination time and sensitivity to involuntary head motion causing artifacts. Extensions of the method to infants will be particularly challenging as this will require RF coils that are more tightly coupled than a standard head coil to achieve better SNR which could be traded for higher spatial resolution needed to account for the smaller dimension of the anatomic structures.

An adult study population was used for this initial investigation due to the larger cranial size of adult participants, ability for adult participants to lay still without sedation during MRI and CT studies, the greater availability of adult volunteers on the authors’ university campus, and the desire to reduce radiation exposure in pediatric patients. We consider the choice of an adult study population reasonable to achieve the investigative goals of this study: to assess agreement between CT and manually segmented MRI and evaluate a multi-atlas segmentation pipeline. The implications of bone-selective MRI affect both adult and pediatric populations. The next step is clearly to assess bone-selective MRI in a pediatric population thus, we are in the process of expanding the current experimental methodology to a cohort of pediatric patients who have already undergone thin-slice CT for clinical purposes. We plan to investigate the accuracy of this technique as a tool for craniosynostosis diagnosis as well as in craniofacial virtual surgical planning. The process of capturing high quality bone-selective MR-images on young children will be facilitated by a newly developed motion corrected imaging pulse sequence (12) utilizing a self-navigated UTE pulse sequence, retrospective motion correction strategy, and k-space mapping approach. The results of this recent study yielded effective removal of motion artifacts and clear depiction of skull bone voxels when performed on head images of four human subjects.

We envision future investigations of both the accuracy of this technology in diagnosing craniosynostosis as compared to the gold-standard, as well as its implementation within virtual surgical planning software at CHOP’s high-volume pediatric center.

CONCLUSION

In the present study, the use of an automated multi-atlas segmentation pipeline eliminates the need of an expert to obtain segmentations. Furthermore, good concordance between the gold standard (thin-slice CT) and the new MRI-based imaging technology was obtained. Bone-selective MRI with automated multi-atlas segmentation has applications in craniofacial surgery as well as extremity surgery, musculoskeletal trauma, and imaging of bone tumors. The elimination of the time for requirements for manual segmentation time should allow bone-selective MRI to be used in clinical practice without causing a delay in treatment even in the case of emergency surgery.

ABBREVIATIONS

DSC

Dice Similarity Coefficient (volume-based measurement used to compare the segmented image with the groundtruth image)

CI

Confidence Interval (defines a range of plausible values for an unknown parameter

HD

Hausdorff Distance (surface-based measurement used to compare the segmented image with the groundtruth image)

RF

Radiofrequency (oscillation rate of electromagnetic field in the frequency range from around 20 kHz to around 300 GHz)

TE

Echo Time (time between the application of the radiofrequency excitation pulse and the peak of the signal induced in the coil)

UTE

Ultrashort Echo Time (shorter echo times than the conventional clinical sequences)

TR

Time of Repetition (amount of between successive pulse sequences applied to the same slice)

VOI

Volume of Interest (specific sub-region in a given image or 3D reconstructed image)

LCC

Lin’s Concordance Correlation (concordance between measurement and a gold standard measurement)

SNR

Signal to Noise Ratio (ratio of a level of desired signal to the level of background noise)