Craniofacial Imaging of Pediatric Patients by Ultrashort Echo-Time Bone-Selective MRI in Comparison to CT

Abstract

Rationale and objectives:

The emergence of low-dose protocols for CT imaging has mitigated pediatric radiation exposure, yet ionizing radiation remains a concern for children with complex craniofacial conditions requiring repeated radiologic monitoring. In this work, the clinical feasibility of an ultrashort echo time (UTE) MRI sequence was investigated in pediatric patients.

Materials and Methods:

Twelve pediatric patients (6 female, age range 8 to 18 years) with various imaging conditions were scanned at 3T using a dual-radiofrequency, dual-echo UTE MRI sequence. Bright-bone images were generated using a weighted least-squares conjugate gradient method to enhance bone specificity. The overlap of the binary skull masks was quantified using the Dice similarity coefficient (DSC) and the 95^th percentile Hausdorff distance (HD95) to evaluate the similarity between MRI and CT. To assess the anatomic accuracy of 3D skull reconstructions, six craniometric distances were recorded and the agreement between MRI- and CT-derived measurements was evaluated using Lin’s concordance correlation coefficient (ρ_c).

Results:

The bright-bone images from UTE MRI demonstrated high bone-contrast, suppression of soft tissue, and separation from air at the sinuses. The DSC and HD95 between MRI and CT had medians of 0.81±0.10 and 1.87±0.32 mm, respectively. There was good agreement between MRI and CT for all craniometric distances (ρ_c ranging from 0.90 to 0.99) with a mean absolute difference in measurements of <2 mm.

Conclusion:

The clinical feasibility of the UTE MRI sequence for craniofacial imaging was demonstrated in a cohort of pediatric patients, showing good agreement with CT in resolving thin bone structures and craniometry.

1. Introduction

Head CT is the most common imaging modality for evaluating craniofacial abnormalities in pediatric patients, accounting for approximately 2.2 million scans per year in the United States¹. Craniofacial surgeons rely on CT as the clinical standard for preoperative and postoperative evaluation of craniofacial skeletal pathologies due to its ability to quickly and accurately capture skeletal morphology². Though technological advances such as low-dose protocols and iterative reconstruction have minimized radiation exposure by CT^{3, 4}, ionizing radiation remains a concern, particularly among children^5–8, who are at higher risk of developing cancer due to increased sensitivity of developing organs and a longer life expectancy during which radiation-related sequelae can manifest¹. For instance, multiple CT scans can triple the risk of leukemia and brain cancer in children⁹. In children who require frequent radiologic assessment, such as those with syndromic craniosynostosis^10–13, fibrous dysplasia¹⁴, and other complex craniofacial conditions, these risks are amplified.

In contrast to CT, MRI does not use ionizing radiation and has emerged as an imaging alternative in recent years. In addition to eliminating radiation exposure, bone-selective MRI has the added appeal of potentially being a “one-stop-shop”, allowing assessment of both bone and soft tissue contrasts in the same scan session, thus aiding in the evaluation of conditions that involve both skeletal and intracranial pathologies, such as craniofacial trauma and craniosynostosis. Several MRI techniques have been developed for skull imaging, the simplest of which was proposed by Eley et al.¹⁵ and termed “black bone” MRI (BB-MRI). This methodology uses a gradient-echo sequence with a low flip angle to obtain proton density-weighted contrast to increase the separation between soft tissue and bone; however, as it relies on the absence of short T₂ signal from bone tissue, air and bone cannot be differentiated at the paranasal sinuses.

Detection of short T₂ protons in osseous tissues requires both the echo time at which k-space center is being acquired and the radiofrequency (RF) pulse to be substantially shorter than T₂ relaxation. Unlike conventional sequences, ultrashort echo (UTE) sequences are solid-state MRI techniques that capture the magnetization of short T₂ signal from collagen-bound water (T₂~0.4–0.5 ms) before it decays¹⁶. One such sequence is the 3D dual-radiofrequency and dual-echo (DURANDE)¹⁷ that utilizes short and long RF pulses to capitalize on the detection sensitivity of bone protons to RF pulse duration and further suppresses soft tissue signal via echo subtraction¹⁸. DURANDE has demonstrated enhanced bone contrast and superior soft tissue suppression compared to other MRI sequences¹⁹, as well as when validated ex vivo and in vivo in healthy adults against CT^{20, 21}.

This study investigates the feasibility of the proposed DURANDE MRI sequence in a cohort of pediatric patients and validates the technique against clinical CT. The MR bone-selective images were evaluated by their ability to 1) resolve thin bone structures from soft tissue and air and 2) quantify the mutual bias of the binary images and craniometric measurements derived from 3D skull renderings compared to those derived from CT.

2. Methods

2.1. MRI Pulse Sequence and Bone-Selective Image Reconstruction

The dual-radiofrequency, dual-echo 3D ultrashort echo pulse sequence (DURANDE) was previously reported in detail (Supplementary Figure S1)¹⁷. Briefly, two RF pulses are applied with different durations and amplitudes along two successive TRs, and within each TR period, two echo images are acquired at two TEs, each sampling different parts of k-space. The short echo of the first RF pulse captures the short T₂ signal from bone (~400 μs), while the longer T₂ signal from soft tissues is retained in all four echoes. View-sharing is used during reconstruction such that echoes at the same TEs are combined for a fully sampled k-space. This obviates the need to acquire two complete datasets (i.e., one with short and the other with long duration RF pulse¹⁸) and hence, reduces the scan time two-fold. The two echo images generated at short and long TEs retain the highest (Image 1, I₁) and lowest bone signal (Image 2, I₂), respectively, and the final bone image (I_bone) is generated by subtracting and normalizing the two magnitude images.

In contrast to the original work in Lee et al.¹⁷, an alternative image reconstruction strategy was implemented in this work as proposed in Vu et al.²² Since the DURANDE sequence interleaves the acquisition of echoes, the two echo images are self-registered with shared edges, resulting in highly correlated spatial positions of the high-magnitude wavelet coefficients. Thus, a joint-ℓ₀ wavelet regularizer in the reconstruction was used to preserve the edge sharpness in the echo subtraction. In previous work^{17, 18}, the normalized subtraction of the two echo images to isolate the short-T₂ species was ill-conditioned because the voxel-wise division amplifies noise anywhere where the proton signal does not exist (i.e., division by near-zero denominator in regions of air voxels in the sinuses and the background). In this work, the bone-selective image is computed by minimizing a weighted linear least-squares objective. Further details are available in the Supplementary Appendix.

2.2. Study Participants

A total of 12 pediatric patients (6 male, median age 14.2 years, range 8.8 to 18 years) were recruited at the Children’s Hospital of Philadelphia (CHOP). Patients indicated for a high-resolution head CT as part of their ongoing medical treatment were considered from February 2023 to February 2024. Eligible participants were children aged 8 to 18 who either had a clinical CT scan performed prior to consent or had a future clinical CT scan scheduled such that CT and MRI scans were no greater than 1 month apart. To avoid sedation procedures, children below the age of 8 were not considered due to concern for their lack of ability to remain still for the duration of the MRI scan. Exclusion criteria included any contraindications for MRI (metallic dental braces, cardiac pacemakers, etc.), pregnancy, and inability to tolerate an MRI scan for 30 minutes without sedation. Written informed consent was obtained from legal guardians of all participants. The study was approved by the authors’ institutional review board and complied with Health Insurance Portability and Accountability Act guidelines.

2.3. Imaging Protocol

All participants underwent a reduced-dose CT scan using a standard clinical pediatric imaging protocol at CHOP. CT images had an in-plane resolution of 0.47–0.53 mm² and 0.75 mm slice thickness. MR images were acquired at a single 3T scanner (Prisma, Siemens, Erlangen, Germany) using a 20-channel head-and-neck coil without sedation. All participants were scanned with DURANDE (Supplementary Figure S1) with the following sequence parameters: FOV of 280 mm³ isotropic, voxel size of 1.1 mm³, TR of 7 ms, TE₁/TE₂ of 0.05/2.45 ms, RF₁/RF₂ duration of 0.04/0.52 ms, flip angle of 12°, dwell time of 4 μs, number of spokes of 50,000, and total scan time of 6 minutes.

2.4. Image Post-Processing

CT images were thresholded to generate bone segmentations and then manually edited to remove the cervical spine and any spurious voxels. The CT images were downsampled to match the MRI voxel resolution and rigidly registered to image I₂ using the mutual information metric via the open-source registration software greedy²³. The registered CT segmentations are considered as the reference ground-truth.

DURANDE images were reconstructed offline using custom code.^{17, 22} Two echo images at short and long TE times were reconstructed, retaining the highest and lowest bone signal, respectively. The bone-selective image is the difference of the two echo images, using the method described in Section 2.1. To segment the bone, a semi-automatic histogram-based method was utilized using custom code in MATLAB (MathWorks, Natick, MA)¹⁹. In brief, the image was segmented using a bone-specific threshold determined from the Gaussian fit of the intensity distribution of the 3D image. The generated mask was modified by a series of morphological operations to remove small, connected components near the edges of the skull and fill in any gaps in the bone mask. To further fine-tune the segmentation and remove erroneous segments, the masks were manually edited using the brush tool in ITK-SNAP version 3.8.0²⁴, which required 1.5 to 3 hours per dataset. The final MRI masks were cropped to include only the cranial vault, orbit, and superior portions of the maxilla.

2.5. Clinician Assessment

A board-certified pediatric neuroradiologist (co-author A.V.) reviewed the CT and MR bone-selective images for all patients. The MR images were evaluated first while being blinded to the CT diagnosis, followed immediately by assessment of the clinical CT images. Using a Likert scale, the reviewer assessed the visualization quality of four sutures including the coronal, sagittal, lambdoid, and occipitomastoid sutures. The Likert scale ranged from 1 to 4, with 1 = poor, 2 = adequate, 3 = good, and 4 = excellent. Half points were given if the suture visualization was partial or asymmetrical. Furthermore, the reviewer assessed the ability to diagnose cranial abnormalities (if present), such as fibrous dysplasia, focal thickening, and glabellar mass. Finally, the reviewer was asked to evaluate both MR and CT images for motion artifacts on a scale of 0 to 2, where 0 = none, 1 = mild, and 2= substantial.

2.6. Data and Statistical Analysis

All MRI and CT segmentations were 3D-rendered in Mimics (Version 23.0, Materialise, Leuven, Belgium). Segmented binary masks were oriented to an axial plane through the superior margins of the bilateral external auditory canals, the left orbitale, and a sagittal plane perpendicular to the axial plane through sella and nasion. Craniometric landmarks were manually identified to measure six anatomic distances: left orbital height, right orbital height, piriform height, cranial width, inter-zygomatic distance, and maxillary width. Measurements were obtained on 3-matic (Version 23.0, Materialise, Leuven, Belgium) using the “measure” tool. Each distance was independently measured three times and the mean value calculated.

The overlap between the MRI and CT segmentations was quantitatively assessed using the Dice similarity coefficient (DSC) and the 95^th percentile Hausdorff distance (HD95). HD95 represents the 95^th percentile of the maximum distances between the surfaces of two objects, making it less susceptible to small outlier subsets because it excludes the largest 5% of values²⁵. Finally, to quantify the agreement in craniometric measurements derived from MRI and CT, Lin’s concordance correlation coefficient (CCC) and Bland-Altman plots were used.

3. Results

Patient characteristics are summarized in Table 1 and Table 2. Twelve children were scanned with both MRI and CT, and none required sedation. CT indications were craniosynostosis follow-up (n=7), glabellar mass evaluation (n=1), subgaleal hematoma drainage follow-up (n=1), fibrous dysplasia (n=1), jaw asymmetry evaluation (n=1), and malocclusion presurgical planning (n=1). Two patients were excluded due to image artifacts from dental braces and excessive motion (Patients 3 and 12 in Table 2), respectively. Thus, 10 patients were included in the final analysis.

Table 1:

Participants’ Demographics.

Cohort Characteristics	N = 12
Age (years) a	13.5 ± 3.0
Sex
Male	6 (50)
Female	6 (50)
Race
White	5 (42)
Black or African American	4 (33)
Asian	1 (8)
Other	2 (17)
Ethnicity
Non-Hispanic or Latino	10 (83)
Hispanic or Latino	2 (17)

^aContinuous variables presented as mean ± standard deviation; categorical data presented as number (percent).

Table 2:

Cohort Characteristics.

Patient No.	Sex	Age (years)α	CT Scan Indication
1	M	16.4	Glabellar mass evaluation
2	F	17.9	Craniosynostosis follow-up
3	M	12.7	Craniosynostosis follow-up
4	M	14.3	Craniosynostosis follow-up
5	M	8.8	Craniosynostosis follow-up
6	F	9.8	Subgaleal hematoma drainage follow-up
7	F	15.7	Fibrous dysplasia evaluation
8	M	12.3	Craniosynostosis follow-up
9	F	9.1	Craniosynostosis follow-up
10	F	15.1	Jaw asymmetry evaluation
11	M	14.0	Craniosynostosis follow-up
12	F	15.8	Malocclusion presurgical planning

^αThe MRI scan was scheduled within a month of CT imaging.

From the DURANDE sequence, two echoes are collected along each TR period with alternating short and long RF pulses. The four resulting echoes then construct two independent k-space datasets via view-sharing echoes at the same TEs. Two sets of images were reconstructed at short and long echo times, from which a bone-specific image was derived using iterative normalized subtraction (Figure 1). The parent technique as described by Johnson et al.¹⁸ and Lee et al.¹⁷ uses a direct normalized subtraction of the two echo images, which results in division by near-zero denominator in regions of air voxels and subsequently leads to false bone signal at bone and air interfaces. These artifacts were eliminated in the iterative normalized subtraction computed with a conjugate gradient algorithm²², enhancing short T₂ specificity and visualization of thin bone structures in the sinuses (Figure 1).

Figure 1:

Examples of reconstructed images from the DURANDE sequence. The first echo image captures the signal from both short and long T₂ species, while second echo image contains only the signal from long T₂ species. The bone-specific image is derived using an iterative normalized subtraction technique computed with a conjugate gradient algorithm²¹. More simply, in previous work, the bone-specific image is the direct normalized subtraction of the two echo images^{16, 17}. The iterative subtraction eliminates noise in air and background voxels and enhances the short-T₂ specificity and visualization of thin bones around the sinuses (yellow arrows).

Example comparisons between CT and MRI are shown in Figure 2 and Figure 3 for four patients evaluated for jaw asymmetry, fibrous dysplasia, craniosynostosis, and subgaleal fluid drainage. The 3D skull renderings of these four patients are depicted in Figure 4 and Figure 5. Like CT, MR images can resolve thin bone structures with superior suppression of soft tissue signal and separation from air in the sinuses. Additionally, suture visualization is observed in both the bone-selective and echo 2 images (Figure 6), the latter having been acquired at a conventional echo time of 2.4 ms yielding a T₁-weighted contrast. As seen in the 2D slices (Figure 3) and 3D renderings (Figure 5), patent osteotomies from prior cranial vault surgeries can also be resolved when validated against CT. The mandible and inferior portions of the maxilla were excluded in the final bone segmentations due to the presence of some motion artifacts (e.g., from mouth breathing), and because the neck channels in the head-neck coil generally have lower SNR.

Figure 2.

MRI and CT images of two participants diagnosed with jaw asymmetry (top) and fibrous dysplasia (bottom), both with no prior head surgery. The contrast of the echo 2 image from DURANDE is that of a conventional T₁-weighted MRI. In the bone-selective MR images, bone has the highest voxel intensity and air appears with background intensity. When validated against CT, MRI can resolve thin bone structures (red arrows), differentiate bone from air at the sinuses (yellow arrows), and visualize some of the calvarial sutures (blue arrows).

Figure 3.

MRI and CT images of two participants. Top: Female patient who underwent fronto-orbital advancement surgery at age 2 for craniosynostosis. The MR images clearly depict patent osteotomies in the parietal bones and vertex (blue arrows). Bottom: Female patient who underwent subgaleal fluid drainage with demonstrated postoperative scalp edema (green arrows). In the bone-selective MR images, bone has the highest voxel intensity and air appears with background intensity.

Figure 4:

MRI and CT skull 3D rendering of two participants with no prior cranial vault surgery (previously shown in Figure 2).

Figure 5:

MRI and CT skull 3D rendering of two participants (previously shown in Figure 3). Top: Female patient who underwent fronto-orbital advancement surgery for craniosynostosis at age 2. The MR images clearly depict patent osteotomies in the parietal bones and vertex. Bottom: Female patient who underwent subgaleal fluid drainage with no history of craniofacial surgery.

Figure 6:

Suture visualization depicted in the UTE MR images compared to CT in two participants (top: 16 yo male; bottom: 15 yo female). The suture details are preserved despite the larger MRI voxel size of 1.1 mm³. The echo 2 image at a conventional echo time of 2.4 ms resembles that of T₁-weighted contrast where there is no signal from bone tissue. The CT images are registered to the MR images to enable direct slice-to-slice comparisons.

The mean absolute differences in craniometric distances between MRI and CT measurements is presented in Table 3. Across all six craniometric variables, the mean absolute percent difference ranged from 0.57% to 1.43%. Agreement between MRI and CT measurements based on Lin’s CCC varied from “moderate” to “almost perfect”²⁶, with ρ_c values ranging from 0.90 to 0.99 across six craniometric variables (Table 3). Bland-Altman plots (Table 3, Figure 7) showed no significant mean bias for craniometric distances between MRI and CT, except for measurements of cranial width (mean bias = −1.2 mm, P=0.011) and interzygomatic distance (mean bias = −1.7 mm, P=0.01) being underestimated in the MRI reconstructions. MRI and CT bone segmentations had a median HD95 of 1.87 ± 0.32 mm (range 1.41 to 2.00 mm). The DSC had a median and standard deviation of 0.81 ± 0.10 (range 0.75 to 0.88) across all participants. Higher DSC values were observed for superior slices, whereas inferior locations (e.g., containing orbits and sinuses) had lower DSC values, contributing to the overall lower average. Small errors in the registration of CT to the MR images may also have exacerbated DSC discrepancies.

Table 3:

Craniometric variables expressed in mean absolute differences between measurements derived from CT and MRI reconstructions. The bias and agreement between the two modalities were assessed with Bland-Altman plots and Lin’s concordance correlation coefficient (CCC).

Craniometric variables	Mean absolute difference (mm)α	Mean absolute difference (%)α	Bland-Altman Mean Bias (mm)	Lin’s CCC (ρc, [95% CI])β
Right Orbit Height	0.42 ± 0.62	1.14 ± 1.65	0.07 (P=0.77)	0.93 [0.76, 0.98]
Left Orbit Height	0.41 ± 0.66	1.13 ± 1.76	0.01 (P=0.96)	0.90 [0.65, 0.97]
Cranial Width	1.25 ± 1.22	0.89 ± 0.85	−1.2 (P=0.01)	0.96 [0.86, 0.99]
Interzygomatic Distance	1.67 ± 1.66	1.35 ± 1.28	−1.7 (P=0.01)	0.93 [0.78, 0.98]
Cranial Length	1.00 ± 1.20	0.57 ± 0.66	−0.63 (P=0.20)	0.99 [0.98, 1.00]
Cranial Height	1.93 ± 2.41	1.43 ± 1.76	1.4 (P=0.15)	0.90 [0.66, 0.97]

^αValues expressed in mean ± standard deviation.

^βLin’s concordance correlation coefficient (ρ_c); 95% confidence interval (CI).

Figure 7:

Bland-Altman plots of the six craniometric measurements to assess bias between MRI and CT across all participants (n=10). The solid lines indicate mean bias and dashed lines represent the limits of agreement (±1.96 SD). There was no significant bias between MRI and CT, except for measurements of cranial width (P=0.011) and interzygomatic distance (P=0.010).

A summary of the radiologist’s assessment of MR and CT images is presented in Table 4. For cranial abnormality detection, all CT images had a score of “excellent”, while the MR images were assessed to have an overall median score of “good” across all patients, such that the images of six patients had scores of “good” to “excellent”, one had an “adequate” score, and three had no cranial abnormalities visible in both CT and MRI. CT and MR images from all 10 patients were evaluated to have no motion artifacts (Table 4, score 0 on a 3-point Likert scale). The coronal and sagittal sutures were not visible in the images of seven and six patients, respectively, in both CT and MRI. For the remaining patients’ MR images, two were scored as “good” (CT scored as “excellent”), and one was scored as “adequate” (CT as “good”) for visualizing the coronal suture. As for the sagittal suture, the MR images of three patients had scores of “adequate” to “excellent” (CT as “excellent”), while one had a score of “poor” (CT as “good”). Moreover, the occipitomastoid suture had a median score of “good”, such that MR images of eight patients had scores of “good” to “excellent” (CT as “excellent”) and two had a score of “poor” (CT scored “adequate” and “excellent”). Finally, the lambdoid suture had a median score of “adequate” where the MR images of four patients had “good” to “excellent” scores (CT scored “good” and “excellent”), three had an “adequate” score (CT scored “good” and “excellent”), one had a “poor” score (CT as “excellent”), and two had no scores due to inability to visualize the suture in both CT and MRI.

Table 4:

Reader assessment scores for suture visualization, abnormality detection, and motion artifacts for the CT and MRI scans.

		Mode Score α		Median Score α
Anatomical Structure	Number of patients assessed	CT	MRI	CT	MRI
Coronal Suture β	3	-	-	4	2.5
Sagittal Suture β	4	-	-	4	2.5
Lambdoid Suture β	8	4	2	4	2.75
Occipitomastoid Synchondrosis	10	4	4	4	3.25
Abnormality detection (if present)γ	7	4	3	4	3
Motion artifacts φ	10	0	0	0	0

^αFour-point Likert scale values of 1 = poor, 2 = adequate, 3 = good, 4 = excellent.

^βSome sutures were not visible in CT and MRI.

^γAbnormality detection for fibrous dysplasia, focal thickening, etc. Three patients did not have visible cranial abnormalities in CT and MRI.

^φMotion artifacts was scored on a three-point Likert Scale of 0 = none, 1 = mild, 2 = substantial.

4. Discussion

CT is the preferred imaging modality for craniofacial pathology due to its superior bone visualization and short scan time. However, the accompanying risks of ionizing radiation may be concerning, particularly for children with complex conditions requiring frequent radiologic monitoring. Bone-selective high-resolution MRI is a promising alternative to CT that eliminates radiation exposure and enables simultaneous characterization of soft tissue and bone. In particular, the DURANDE sequence has demonstrated superior bone contrast and soft tissue suppression compared to other bone-selective MRI sequences in the healthy adult population¹⁹. However, this technique has not previously been validated among pediatric patients—the population that is projected to benefit most from a non-radiative alternative^5–8. In this study, the clinical feasibility of the proposed MRI sequence was evaluated against CT in a cohort of pediatric patients and their agreement was quantitatively assessed. The MR images depicted high bone contrast, particularly for thin bone structures, suppression of soft tissue, and separation of bone from air in the sinuses. The findings show good agreement in all craniometric variables (ρ_c≥0.90) measured from 3D skull reconstructions derived from CT and MRI. Furthermore, the radiologic assessment demonstrated the ability to detect sutures and cranial abnormalities from the bone-selective MR images in most patients. The discrepancy between MRI and CT is attributed to the larger voxel size of MRI (1.1 mm³)

Six craniometric measurements were used to quantitatively evaluate the agreement in 3D skull reconstructions between MRI and CT. Measurements across all patients had a strength-of-agreement ranging from “moderate” to “almost perfect”²⁶ and no mean bias, aside from cranial width and interzygomatic distance. This discrepancy may be due to the small sample size, intra-rater variations, and thinner bone anatomy of the zygomatic arch, exacerbated by the larger voxel size of MRI (1.1 mm³). Nonetheless, mean bias in measurements were 1.25 mm and 1.67 mm for cranial width and interzygomatic distance, respectively, while HD95 was less than 2 mm when assessing the boundary overlap of MRI and CT skull segmentations. Previous studies have found intra-rater and inter-rater variability for cranial anthropometric measurements to be within 2 mm^27–32.

MRI is a safe option for brain and skull imaging in pediatric patients. The simplest of these MRI-based approaches, BB-MRI, uses a 3D gradient echo sequence with a low flip angle to acquire proton density-weighted soft tissue contrast^{15, 33, 34}. This technique relies on the absence of bone signal to separate bone and soft tissue voxels. In recent years, there has been a growing interest in the clinical potential of MRI for craniofacial imaging; for instance, Low et al. outlined a BB-MRI protocol implemented at the authors’ institution as an alternative to CT for evaluating pediatric patients with traumatic injuries, fibrous dysplasia, and craniosynostosis³⁵. Another group used a BB-MRI technique with a fast low-angle shot golden-angle 3D stack-of-stars radial VIBE sequence for cranial bone imaging, demonstrating high sensitivity and specificity for calvarial suture closure and fracture detection in children^36–38. Other studies have assessed the utility of BB-MRI in children with posterior plagiocephaly^{39, 40}, metopic craniosynostosis^{34, 41}, and sagittal craniosynostosis^{34, 40}. However, the main limitation of BB-MRI is that it cannot differentiate between air and bone at the sinuses, which limits its utility for evaluating lesions at air-bone interfaces³⁵. This constraint complicates skull tissue segmentation and the subsequent tasks of 3D rendering and surgical planning.

Unlike BB-MRI, UTE and zero echo time (ZTE) sequences capture the short T₂ signal from osseous tissues, thus discerning air and bone^{42, 43}. ZTE methods use low flip angles to yield proton density-weighted images which contain nonzero signal from both bone and soft tissue. Lu et al. demonstrated that ZTE and clinical CT provide qualitatively similar bone detail in pediatric patients with various diagnostic indications, including craniosynostosis and head trauma⁴⁴. In another study, ZTE MRI was used to assess skull fractures in adult patients showing substantial interobserver agreement for overall image quality when validated against CT⁴⁵. Kralik et al. utilized a UTE sequence with pointwise encoding time reduction with radial acquisition for imaging head trauma in children, showing good sensitivity and high specificity for the diagnosis of skull fractures⁴⁶.

There are several benefits to the proposed DURANDE sequence in this study compared to previously published techniques. The dual RF design of alternating short and long pulses was previously shown to enhance bone selectivity in the echo subtraction image compared to standard UTE sequences¹⁸. This is because bone proton magnetization has a short T₂ relaxation time, making it sensitive to the duration of the RF pulse, while soft tissue protons maintain almost the same signal intensity in each acquired echo. As a result, the bone signal is maximized in the echo subtraction image. In addition to the bone-specific image, the proposed UTE sequence yields two co-registered soft tissue T₁-weighted contrasts, which may make the sequence useful for assessing both bone and soft tissue contrasts within the same scan. Because of the echo subtraction approach, the bone-selective image from DURANDE is self-normalized and does not require bias-field correction, simplifying the post-processing pipeline. While ZTE methods provide significantly improved bone specificity over BB-MRI, the separation of bone, soft tissue, and air signals is achieved by applying bias-field correction, inverse logarithmic scaling, and histogram-based thresholding⁴². This approach has limitations since there is overlap in grayscale intensity distributions between bone and air voxels as well as between bone and soft tissue voxels, resulting in false-positives (e.g., other short T₂ species) and false-negatives (e.g., unmineralized and thin bones)^{19, 43}. It is noted that even UTE sequences can yield false-bone signal in the temporalis tendon due to its short T₂¹⁹. Furthermore, for the short T₁ of cortical bone water (~400 μs ¹⁶), a larger flip angle is needed to maximize the signal from bone tissue. ZTE techniques have inherently low signal-to-noise ratio (SNR) due to the small flip angle required to achieve 3D proton-density weighted contrast with zero echo time and short repetition times^{42, 43}, unlike UTE sequences that can incorporate higher flip angles, yielding higher SNR. The implementation of bone-selective MRI in clinical practice offers several benefits to CT imaging. First, the burden of ionizing radiation is eliminated, particularly for those with conditions requiring frequent radiologic imaging, such as children with fibrous dysplasia and syndromic craniosynostosis. Second, bone-selective MRI can offer a “one-stop-shop” with high-resolution characterization of both soft tissue and bone for patients with both soft tissue and skeletal dysmorphology. Beyond the skull, this technique may also be applied to the remaining axial and appendicular skeleton. Finally, large-scale implementation of a bone-selective technique supplementing traditional MRI will enable the acquisition of a high volume of scans with “normal” skeletal morphology to serve as a control cohort in craniofacial research⁴⁷.

Several limitations of this study should be noted. First, motion-artifacts were present in one patient that was excluded in this study. Without sedation, children may have difficulty remaining still for the duration of the scan. If sedation is required, the risk-benefit-ratio of ionizing radiation from CT versus anesthesia risks of undergoing sedation for MRI must be assessed. Reducing scan time comes at the expense of acquiring fewer k-space views, resulting in lower SNR. However, shorter scan times can now be achieved with the recent incorporation of parallel imaging and compressed sensing techniques for bone-selective craniofacial imaging²². Second, unlike the submillimeter resolution in CT, the voxel resolution in this MRI protocol was 1.1 mm³, which may be inadequate for younger children with smaller craniofacial structures. Therefore, there is a need for a higher-resolution MRI protocol for imaging pediatric patients, coupled with a smaller head-and-neck coil to compensate for some of the compromised SNR. Third, the semi-automatic segmentation approach used in this study required manual editing of the final bone masks, which is a time-consuming task requiring specialized knowledge of skull anatomy. Assuming three voxel signal compartments comprising bone, soft tissue, and air, the distribution of grayscale intensities in the MR images overlap more than the Hounsfield units in the CT images, contributing to the difficulty of segmenting bone in MRI. Additionally, the segmentations did not include the mandible or inferior portion of the maxilla. An automated deep-learning approach for full skull segmentations has the potential to provide a fast and robust segmentation pipeline for clinical translation of the proposed technique. This method was not achieved in this work due to the limited cohort size and the heterogeneity of skull phenotypes that were insufficient for training a deep-learning model.

Conclusion

In summary, the clinical feasibility of the DURANDE MRI technique for craniofacial imaging was investigated in a cohort of pediatric patients and validated against clinical CT. The MRI technique can clearly depict thin anatomical structures with high bone contrast, suppression of soft tissue signal, and separation from air at the sinuses. There was good agreement in all craniometric variables measured from 3D skull renderings derived from CT and MRI scans. Future studies will evaluate the proposed UTE technique in a larger cohort for the preoperative assessment of younger children.

Abbreviations List:

UTE

Ultrashort echo time

DURANDE

Dual-radiofrequency and Dual Echo UTE

ZTE

Zero echo time

BB-MRI

Black-bone MRI

DSC

Dice similarity coefficient

HD95

Hausdorff distance 95^th percentile

CCC

Concordance correlation coefficient