Skip to main content
NCBI home page
Oral Health & Preventive Dentistry logoLink to Oral Health & Preventive Dentistry
. 2025 Nov 26;23:701–710. doi: 10.3290/j.ohpd.c_2365

Geometric Fidelity of Magnetic Resonance Imaging and Computed Tomography-Derived Virtual 3D Models of Porcine Cadaver Mandibles: Conventional Versus Artificial Intelligence-Based Segmentation

Lucas M Ritschl a, Katharina Pippich b, Matthias Herrmann c, Herbert Deppe d, Anton Sculean e, Monika Probst f, Florian A Probst g
PMCID: PMC12650764  PMID: 41294019

Abstract

Purpose

The workflow for virtual surgical planning (VSP) and the application of CAD/CAM (computer-aided design/computer-aided manufacturing) procedures are mainly based on computed tomography (CT) derived DICOM data sets. Alternatively, this study aims to preclinically illuminate the feasibility of a magnetic resonance imaging (MRI) based workflow and the impact of artificial intelligence (AI) based segmentation on the required fidelity on basic 3D geometry acquisition.

Materials and Methods

Porcine cadaver mandibles were imaged with CT and a T1-weighted MRI sequence. The resulting DICOM data sets were segmented conventionally (Mimics Medical 17.0, Materialize; Belgium) and with AI-based segmentation software (ImFusion Labels and Suite, Version 2.19.2, ImFusion; Germany). The four standard tessellation language (STL) files were superimposed with a corresponding reference model derived from an optic scan (Artec Space Spider, Artec 3D; Luxembourg) and the following parameters were analysed: Hausdorff distance (HD), mean surface distance (MSD), root mean square distance (RMSD), time.

Results

In comparison to the reference model, all four parameters were significantly (P <0.001) better for the CT imaging and the AI-based segmentation. MRI-derived AI-based segmentation reached the fidelity of CT imaging data sets and conventional segmentation (HD, MSD, and RMSD each P >0.05).

Conclusion

The use of AI-based segmentation software proved to be useful and feasible for MRI-derived data sets, and generated the desired 3D geometry more quickly while maintaining the necessary quality. Nevertheless, the results for the CT were still better and remain yet the standard.

Keywords: artificial intelligence, computed tomography, computer-assisted surgery, image processing, magnetic resonance imaging, segmentation

Virtual surgical planning (VSP) and the application of CAD/CAM (computer-aided design/computer-aided manufacturing) technology are nowadays the gold standard in cranio-maxillofacial surgery, including reconstructive surgery, traumatology, implantology, and orthognathic surgery. With special regard to reconstructive microsurgery, the application offers for instance shorter surgery and ischaemia times and hospital stays with improved symmetry, aesthetics, bony consolidation rates, and function.15,31,42,43 In addition, it was also shown that virtual planning can be reliably translated into the operating room.28 But according to the commonly applied digital workflow sequence, the planning is usually based on a digital images and communications in medicine (DICOM) data set of a computed tomography (CT) or cone beam computed tomography (CBCT) scan. As a result, every patient who undergoes CAD/CAM-assisted surgery is currently potentially exposed to additional radiation. However, according to the ALARA principle (as low as reasonably achievable), radiation exposure should be reduced to the lowest possible level in order to keep the risk of radiation-associated consequences as low as possible.

Thus, with regard to facial bony reconstruction, especially of the mandible with alloplastic materials or microvascular bone flaps – using the free fibula, iliac crest, or scapula – radiation exposure is currently inevitable. Despite constant efforts to further reduce the radiation dose,33,50 there is a general desire for an alternative solution to obtain a DICOM data set for further processing in the digital workflow – the segmentation and thus the conversion of the DICOM data set into, for example, a standard tessellation language (STL) file. However, the necessary accuracy in the acquisition of the geometry should be ensured within a reasonable time frame. On this topic, it has already been described by others that linear measurements between MRI- and CT-derived 3D surface models correspond well.6,10,14,19 Additionally, our research group has recently investigated the possibility of adequately capturing the complex mandibular geometry using a T1-weighted magnetic resonance imaging (MRI) sequence compared to CT or CBCT on the porcine mandible.24 The results of this study indicated that the accuracy and reliability of T1-weighted MRI-derived virtual 3D bone surface models were equal to CT and CBCT. But segmentation proved tedious in parts and also took longer than for CT- or CBCT-derived data sets. Consequently, with this follow-up study, we now aimed to investigate the impact of artificial intelligence (AI) based segmentation on accuracy and time required to make the promising use of MRI-based VSP in cranio-maxillofacial surgery more feasible.

MATERIALS AND METHODS

Ethical Statement and Animal Welfare

The prospective study was planned and performed according to the STARD guidelines. The Municipal Veterinary Office (city of Munich, Germany) approved the cadaveric study. It was not necessary to obtain approval from the institutional ethics committee of the Technical University of Munich, Klinikum rechts der Isar, because this study was a porcine cadaver study of animals already slaughtered independently of this study.

Study Design

This study was planned and performed using the CT- and MRI-derived DICOM data sets, as well as the acquired optical scans of the preliminary study.24 But, following segmentation, alignment and analyses were performed completely anew and independently from the preliminary study. The surrounding soft tissue was not removed before imaging (CT or MRI).24 One part of the mandibular body (n = 10) served as the region of interest (ROI). The segmentation was done either with a semi-automatic algorithm or an AI-based algorithm derived from CT and MRI images. The segmentation processes were carried out independently by two observers twice (repetition after 6 weeks) for each segmentation algorithm and imaging modality. The resulting mandibular 3D surface geometries were exported as STL files and then compared with corresponding optical scans (= ground truth model) by digitally superimposing them via an alignment procedure (Fig 1).

Fig 1.

Fig 1

Open in a new tab

Study design.

In order to also examine the time component of the segmentation, the time required for segmentation of the respective lower jaw halves, and to convert the given DICOM data set into an STL file, was recorded.

Imaging – computed tomography

The porcine cadaver mandible was aligned with the occlusal plane horizontally to the ground. Then, multi-slice CT images (Philips Ingenuity 128 device, Philips Healthcare; Best, Netherlands) were created with the following parameters: isotropic voxel size 0.67 mm; field of view 20–20 cm; tube voltage 120 kV; tube current-time product 250 mAs.

Imaging – magnetic resonance imaging

The porcine cadaver mandibles were placed in a tub that was filled with water. The occlusal plane was aligned approximately horizontally to the floor, and a 16-channel head and neck spine array was placed around the box. The three-dimensional T1-weighted MRI sequence was generated using a 3Tesla MRI system (Elition, Philips Healthcare; Best, Netherlands). This sequence was optimised for bone visualisation and is already in clinical use.25 In order to reduce the echo time, partial Fourier imaging in the direction of frequency coding with a factor of 60% was used.12 The following parameters were applied to the recordings: Acquisition time 03:08 min; field of view 180 mm; matrix 420 × 419; recorded voxel size 0.6 × 0.6 × 0.6 mm3; number of signal averages 1; repetition time 10 ms; echo time 1.55 ms; compressed sensing (CS) + sensitivity encoding (SENSE): yes; reduction factor 2.3.

Optical scan of bone surface

After the CT and MRI images were taken, all soft tissue was removed from the porcine cadaver mandibles in order to create an optical three-dimensional scan of the lower jaw (Artec Space Spider, Artec 3D; Luxembourg), which served as a reference (= ground truth model) 47. The cadavers were only processed with conventional dissecting instruments. Neither thermal nor chemical treatment was applied to the bone in order to avoid possible damage to the bone surface and not to falsify the result of the geometric surface comparisons.11,45 The manufacturer of the Artec Space Spider scanner specifies a resolution of up to 0.1 mm and a point accuracy of up to 0.05 mm. With the help of the appropriate software (Artec Studio 13, Artec 3D; Luxembourg) and the associated autopilot tool, the data generated in this way were ultimately used to generate three-dimensional scans. In order to use these as a reference and to be able to compare them with the segmentation results, they were exported as an STL file.24 The Artec Studio software provides a low maximum error value of 0.5 mm for post-processing scans after fine registration, alignment and global registration.

Image processing and segmentation procedures

Semi-automatic/conventional algorithm

The conventional semi-automatic algorithm for segmentation was carried out with Mimics (Mimics Medical 17.0, Materialize; Belgium). After opening the Mimics Medical program, the CT- or MRI-derived DICOM data set of the desired porcine cadaver mandible was opened and positioned in the coordinate system. The time measurement was started immediately afterwards.

The thresholding function was used to segment the CT images (Fig 2), applying the predefined ‘Bone CT’ classification (all voxels with Hounsfield units of 226 and higher). Then the ROI was determined.

Fig 2a and b.

Fig 2a and b

Open in a new tab

Porcine cadaver mandible after (a) ‘dynamic region growing’ (axial, MRI) and (b) ‘thresholding’ (axial, CT).

The ‘Dynamic Region Growing’ function was used to segment the MRI images. Therefore, all points (seed points) on the jaw were defined, which had assumed values between 0 and 10 GV (grayscale value). The range of ‘Dynamic Region Growing’ was also set between 40 and 50 GV (difference to seed point value) in each case individually. The result of the ‘Dynamic Region Growing’ function is shown in Figure 2.

After thresholding or region growing, manual post-processing of the labels was still necessary for precise segmentation. For this reason, with the help of the ‘Multiple Slice Edit’ tool and the integrated interpolation function, the residual holes and incorrectly marked areas were inspected and repaired manually. Immediately afterwards, the segmented geometry was saved and exported as an STL file, and the time measurement was stopped and noted.

Automatic/AI-based algorithm

The AI-based, automatic segmentation was done with the ImFusion Labels and Suite software (Version 2.19.2, ImFusion; Germany). These coherent programs allow the built-in AI to almost completely dispense with any manual post-processing. By appropriately setting a few seed points on individual sectional images of the processed modality, which tell the software where the tissues to be kept apart are located, for example, bone and soft tissue, the automatic algorithm segments the DICOM data across all sectional images and, after a brief check by the observer, converts it into an STL file. The program thus not only provides interpolation between processed sectional images, but also calculates the complete 3D model itself based on the sectional image, without any edge drawing at the tissue boundaries, with only minimal input from the observer.

The following descriptions applied equally to both modalities. After importing and opening the desired DICOM data set, the ImFusion Labels software switched to annotation mode, and time measurement was started. At the beginning of segmentation with this software, no thresholding or region growing was used. The ROI was defined, whereupon the segmentation was started by selecting the ‘Interactive Segmentation’ button.

The settings have now been made on the cursor, which was the main tool for segmentation. First, the size was set to a range between 6 and 20 mm. The second adjustable option was the adaptiveness of the cursor, whereby adaptiveness defined how much the cursor adjusted to the objects. The higher the setting, the more likely it was to protect against drawing beyond the object. Specifically, this meant that while the cursor was being moved over the image, it was shown where the program distinguished the bone from the soft tissue or from the background. However, if the adaptiveness was too high, an overcorrection would result. This overcorrection could largely be avoided with an adaptiveness value between 0.3 and 0.7.

In order to begin the desired segmentation, background label seeds were first set (Fig 3a). This was set on a total of five to ten sectional images in the axial section. Then foreground label seeds were placed on the mandibular bone (Fig 3b), also on five to ten layers. Based on these seeds, the software automatically calculated the foreground label completely. Only individual parts of the label had to be corrected by adding additional background and foreground seeds.

Fig 3a and b.

Fig 3a and b

Open in a new tab

ImFusion Labels in axial layers (a) setting of the background markers and (b) setting of the foreground markers.

After the result was accepted, the program was redirected to ImFusion Suite. After visual inspection of the created 3D geometry, the result was saved and exported as an STL file, and time measurement was stopped and noted.

Alignment and Cropping

Following, the exported CT- or MRI-derived STL files were aligned and compared to the corresponding optical 3D ground truth model using the software Artec Studio Professional 13 (Artec 3D; Luxembourg). For this purpose, the corresponding STL file was imported and superimposed using the ‘Align’ function. A rigid ICP (iterative closest point) alignment algorithm was employed. The manual rigid alignment procedure was used for this purpose, which only allowed rigid, linear movements and rotations when the objects are superimposed without changes in scale or similar settings 30, based on a three-point alignment. Hence, three corresponding points were selected that were clearly visible and reproducible: the most occlusal tip of the second premolar, the most distal tip of the second molar, and the mental foramen. Based on this manual orientation, the geometries were finally automatically aligned by the software.

To minimise an error in the superimposition measurements, the previously overlaid geometries were ‘cropped’ (trimmed) in Blender (Version 2.82a) to a matching dimension in all three planes. This was to avoid incorrectly high results by measuring incorrect pairs of correlated points with each other.

Calculation of Geometric Deviations

Based on the alignments, the surface differences between the CT- or MRI-derived STL files and the corresponding optical 3D scans were determined. In order to quantify the differences, we have included key metrics for assessing geometric fidelity, such as the Hausdorff distance (HD),2,27,49 mean surface distance (MSD)30,47 and root mean square distance (RMSD).11,29,30,36 In particular, Hausdorff distance is a well-established method for quantifying the maximum geometric deviation between two-point sets. It considers the greatest deviation within the entire point cloud, rather than just isolated point-pair differences. This provides a more robust and comprehensive analysis of segmentation accuracy, which is clinically more relevant in the setting of CAD/CAM surgical planning. Additionally, pairwise point comparisons offer only fragmented information about the overall geometry of the models.

These measurements were carried out by the open-source software MeshLab (Version 2021.07). In addition, colour maps were used to visualise any geometric deviations (Fig 4).

Fig 4.

Fig 4

Open in a new tab

Colour map of the surface deviation measurements [mm] of the segmented porcine cadaver mandibles compared to the optical 3D scan (histogram showing the distributions of the deviations of the corresponding points in mm: red colour indicating higher deviation; blue colour indicating lower deviation; MeshLab (Version 2021.07)).

Statistics

The intraclass correlation (ICC) coefficient was calculated to determine the intra- and interrater reliability and consistency of measurements performed by two raters applying a two-way mixed model (MH and KP). This was calculated based on the RMSD. As the primary outcome variable, three-dimensional geometric deviations between the given segmentation algorithms or the modalities and the optical 3D scans were set, given as the HD, the MSD, and the RMSD. The time required for segmentation of one half of the jaw based on either algorithm and modality was determined as the secondary outcome variable. Continuous measurements are presented as median (range). In not normally distributed data sets, the Wilcoxon rank sum test was used for dependent samples. All statistical tests were performed on an exploratory two-sided 5% significance level. No adjustments were made for multiple testing. Bland–Altman analyses were performed to additionally assess the agreement of the combinations of the applied segmentation algorithms and modalities with the most common combination, the semi-automatic segmentation of CT data sets. These show their mean differences (MSD values) and upper and lower limits of agreement (mean value of the differences ±1.96 × standard deviation of the differences). Analysis was performed using Excel for Mac software (Microsoft, Redmond, USA) and IBM SPSS 24 for Mac software (IBM, Armonk; New York, USA).

RESULTS

For the interrater reliability, the ICCs, based on the RMSD, were 0.955 (CT), 0.915 (MRI), 0.933 (semi-automatic), and 0.948 (automatic). Intrarater ICCs were 0.943/0.974 (CT), 0.909/0.928 (MRI), 0.915/0.956 (semi-automatic), and 0.951/0.953 (automatic).

The median geometric deviations of the segmentation results compared to the optical scan were lower for the CT-derived data sets and the automatic segmentation algorithm and are shown in Table 1. These results were highly significant (P < 0.001) for all comparative parameters examined (HD, MSD, and RMSD). A similar result was found for the segmentation time. It was significantly (P <0.001) shorter with CT in comparison to MRI, and with the automatic algorithm in comparison to the semi-automatic algorithm.

Table 1.

Comparison of modalities and algorithms

Modality

Algorithm

CT

MRI

Semi-automatic

Automatic

HD [mm]

Count

80

80

80

80

Median

5.282

5.732

5.736

5.153

Minimum

3.117

3.761

3.674

3.117

Maximum

7.308

7.894

7.894

7.304

MSD [mm]

Count

80

80

80

80

Median

0.412

0.466

0.502

0.402

Minimum

0.308

0.365

0.365

0.308

Maximum

0.628

0.752

0.752

0.527

RMSD [mm]

Count

80

80

80

80

Median

0.582

0.679

0.663

0.576

Minimum

0.418

0.515

0.472

0.418

Maximum

0.744

0.851

0.851

0.800

Time [s]

Count

80

80

80

80

Median

848

1504

1321

939

Minimum

505

835

685

505

Maximum

1128

2363

2363

1493
Open in a new tab

Combination of Modality and Algorithm

More precisely, the modalities were combined with the segmentation algorithms and compared to standardised procedure: CT imaging with subsequent semi-automatic segmentation.

Median geometric deviations of the HD between the segmentation results and the optical scans were 5.746 (3.674– 7.308) mm for the combination of CT/semi-automatic algorithm, 5.732 mm (4.311–7.894) for MRI/semi-automatic algorithm, 4.656 mm (3.117– 5.810) for CT/automatic algorithm, and 5.746 mm (3.761–7.304) for MRI/automatic algorithm. Only the differences between CT/semi-automatic algorithm and CT/automatic algorithm were statistically significant (P <0.001) (Fig 5a).

Fig 5a to d.

Fig 5a to d

Open in a new tab

Comparison of different parameters depending on the used modality and the method of segmentation: (a) Hausdorff distances (HD), (b) mean square distances (MSD), (c) root mean square distance (RMSD), and (d) time.

Median geometric deviations of the MSD between the segmentation results and the optical scans were 0.465 mm (0.365–0.628) for CT/semi-automatic algorithm, 0.569 (0.406–0.752) mm for MRI/semi-automatic algorithm, 0.385 mm (0.308–0.527) for CT/automatic algorithm, and 0.425 mm (0.365–0.520) for MRI/automatic algorithm. Differences between CT/semi-automatic algorithm, MRI/semi-automatic algorithm, and CT/automatic algorithm were all statistically highly significant (P <0.001). Differences between CT/semi-automatic algorithm and MRI/automatic algorithm were also significant (P <0.032) (Fig 5b).

Median geometric deviations with reference to RMSD between the segmentation results and the optical scans were 0.623 mm (0.472–0.744) for CT/semi-automatic algorithm, 0.718 mm (0.565–0.851) for MRI/semi-automatic algorithm, 0.521 mm (0.418–0.656) for CT/automatic algorithm, and 0.629 mm (0.515–0.800) for MRI/automatic algorithm. Differences between CT/semi-automatic algorithm, MRI/semi-automatic algorithm, and CT/automatic algorithm were all statistically highly significant (P <0.001). Differences between CT/semi-automatic algorithm and MRI/automatic algorithm were not significant (P <0.340) (Fig 5c).

Median segmentation times of the examined modalities/algorithms were 932 (685–1128) s for CT/semi-automatic algorithm, 1906 s (1514–2363) for MRI/semi-automatic algorithm, 767 s (505–955) for CT/automatic algorithm, and 1276 s (835–1493) for MRI/automatic algorithm. All time differences based on the combination of the imaging modalities and the segmentation algorithms used, compared with the CT/semi-automatic algorithm, were highly significant (P <0.001) (Fig 5d).

The Bland–Altman analysis indicated high agreement between the compared combinations of the applied segmentation algorithms and modalities with the most common combination, the semi-automatic segmentation of CT data sets, based on the MSD, with 92.5% (37 of 40 differences of the compared measured values of each of the three Bland–Altman plots) within the limits of agreement (Fig 6).

Fig 6.

Fig 6

Open in a new tab

Bland–Altman plots of the differences between the compared combinations of applied segmentation algorithms/modalities and the semi-automatic segmentation of CT data sets with regard to mean surface distances (MSD).

DISCUSSION

This study highlights two important and interesting aspects. First, T1-weighted MRI-derived data sets enable sufficiently accurate 3D surface models of porcine cadaver models in comparison to a ground truth model. Second, the time-consuming procedure of bone segmentation in MRI-derived images was significantly reduced by applying AI-based segmentation software, while preserving the required overall accuracy. According to the evaluated 3D parameters, this reached the level of known and well-accepted accuracy as it is seen in the commonly applied workflow of CT-derived DICOM data sets and use of licence-based semi-automatic segmentation software, for example, Mimics software. When considering the automatic algorithm, the algorithm examined in this study is not a fully automatic process, as has been tested in other studies.44 The applied segmentation algorithm did not need to be trained with preceding data and could therefore be used immediately. Only the already described interactive contributions of the user were needed to let the automatic part run. This had the added advantage that the user himself had some decision-making confidence, but the work could be significantly reduced by ImFusion’s algorithm. Also, although there is little research on the segmentation algorithm of ImFusion or comparable programs based on random walker and watershed segmentation, this was shown in a study by McGrath et al,22 but related to other anatomical structures. There, it was described that ImFusion’s algorithm was less time-consuming and physically demanding to use, with the accuracy additionally being improved, compared to manual segmentation. Both advantages could be confirmed in our study with regard to mandibular segmentation.

The focus of MRI imaging was mainly on soft tissue for a long time. In the precise 3D assessment of bony structures (such as the mandibular canal), CT or CBCT still play a primary role. The introduction of new sequences enabled reasonable MRI-based judgement in musculoskeletal imaging including the quantification of cortical and cancellous bone.7, 21, 35,37 Consequently, it has become increasingly recognised in the workflow of VSP and in CAD/CAM procedures in recent years as an alternative 3D cross-sectional imaging modality to CT and CBCT, which is not associated with ionising radiation. As previously reported, 3D surface model generation of our T1-weighted MRI sequence revealed comparable accuracy to CT- or CBCT-derived data sets.24 In this follow-up study we have included the HD as an additional, critical 3D parameter and compared the segmentation results of two different software solutions. The new insights are of clinical relevance, as the majority of the CAD/CAM workflow for cranio-facial surgery is based on radiation-associated CT- or CBCT-derived data sets. These surgeries also include elective procedures – like orthognathic surgery and dental implant insertion – and procedures in infants – like the treatment of craniosynostosis 25, 40. Patients undergoing these CAD/CAM-assisted procedures are consequently at risk of increased ‘unnecessary’ radiation exposure, although, on the other hand, postoperative results improve by including VSP. Eley et al reported on a novel concept in MRI sequencing to enhance the bone–soft tissue boundary by minimising soft tissue contrast, calling this sequence ‘black bone’.7–9 Saarikko et al described their first impressions in preoperative evaluation of patients with craniosynostosis comparing CT with their black bone MRI sequence.32 In their study, black bone MRI imaging was equivalent for the evaluation of cranial sutures and intracranial impression, but additionally enabled the assessment of the brain structures in one imaging session. Sychuta et al investigated the accuracy of black bone MRI-derived surgical planning and 3D cutting guide creation for cranial vault remodelling compared with a conventional CT-based workflow.40 In this study, comprising ten fresh human cadaver head specimens, they showed that postoperative reconstructions deviated by less than 1.5 mm from the preoperative plan, independently of CT- or MRI-based imaging. They thus demonstrated in a preclinical setting that black bone MRI imaging may be applied for CAD/CAM-assisted calvarial vault surgery with comparable results to the current standard of care.40

Virtual surgical planning is also frequently used in mandibular or maxillary reconstruction. Different CAD/CAM solutions have been reported, which are mainly based on radiation-associated imaging modalities.28 Sychuta et al also analysed their black bone MRI sequence for mandibular reconstruction with free fibula flap fresh human cadaver specimens. They were able to demonstrate the feasibility of MRI-based VSP and surgical accuracy comparable to CT-based VSP.41

Inaccuracies may occur at any step of VSP, but are predominantly attributed to imaging and image processing, including the critical step of segmentation.18 Manual segmentation or global thresholding influences the result of segmentation significantly.46 Automatic, AI-based solutions may decrease time and inaccuracies, as shown in this study. The AI-based application shown here offered the obvious advantage of saving human resources. On the one hand, this saves time, since the observer only initiated the segmentation and was thus liberated from any tedious edge processing. Consequently, considerably fewer incorrectly marked areas had to be fixed. On the other hand, this circumstance also had an impact on the accuracy of the results, since each additional manual step, which was often necessary with the semi-automatic algorithm, is a potential source of error.

Still limited and heterogenous data are available assessing the trueness, precision and accuracy of MRI-derived 3D bone surface models. We have analysed the fidelity of the 3D surface models in a structured and sound way. ICC coefficients of all 3D parameters (RMSD, MSD, and HD) consistently showed very good intra- and interrater reliability (ICC >0.9). This makes our results relevant and allows further discussion of them. But the comparison of 3D surface models remains difficult, as different parameters and measurements are described in the literature to define ‘accuracy’ in the new digital 3D era. For example, Barr et al were not able to specify a valuable parameter and result for ‘accuracy’ in a recent systematic review and meta-analysis on VSP in mandibular reconstruction.3 The reason for this is the heterogeneity of parameters, as the majority of studies compared only 2D linear or angular parameters for fibula segment lengths,13,16,34,39 intercondylar distance,23 or by comparing interfragmentary gap distances.38 In our opinion, this is no longer up to date, which is why our focus in this study has been on pure 3D parameters that encompass the entire porcine cadaver mandibular geometry. This approach seems to be a more contemporary and objective way to compare two similar and relatively complex objects. Within the applied parameters (HD, MSD, and RMSD), HD is the most critical one, as it defines and highlights the worst superimposition of corresponding points of the whole geometry, and RMSD is the most commonly applied in recent literature. In this study, the combination of CT imaging and AI-based segmentation resulted in the best HD (Fig 5a). The RMSD showed for all combinations of imaging modality and segmentation software measurements below 1 mm, which is supposed to be clinically relevant.

Regardless of the fact that MRI protects patients from radiation exposure, it has diagnostic added value compared to CT and CBCT due to its excellent soft tissue depiction, especially in neurocranium imaging or in the staging of soft tissue tumours.20 Accordingly, relevant parts of the dental and facial anatomy, such as dental pulp and mucosa or neurovascular structures, skin, and subcutaneous tissues, may be better integrated into the virtual planning process.1,4,5,17,25,26 Future studies will need to show how other software solutions, including open-source software, enable feasible and reliable MRI-derived bone models.48

Limitations

First, the study design was based on a purely cadaveric study and may not display and represent results for human mandibles, for example. Nevertheless, the authors believe that this study will add interesting information. Further, comparisons on human mandibles are planned. Second, the used AI-based software is a licence-based solution and may not be available for all potential users.

At the time of the study, only the hand-held scanner with the above-mentioned point accuracy of 50 micrometres was available to the research group. This can already cause a bias in the results, as there are more accurate scanners on the market today. However, this accuracy is sufficient for the segmentation of mandibles, for example, to produce saw templates. Nevertheless, the results presented here reflect an important trend that needs to be followed in the future.

CONCLUSION

Segmentation of MRI-derived DICOM data sets using a bone-optimised T1-weighted MRI sequence of porcine cadaver mandibles was feasible and generated 3D surface models with comparable accuracy to the corresponding conventional CT-derived DICOM data sets. The use of an AI-based software solution proved to be useful and reliable, and created the desired 3D geometry more quickly while maintaining the necessary quality. CAD/CAM procedures may be planned on the basis of MRI imaging more frequently in the future.

Acknowledgement

The authors did not receive financial or material support. MRI examinations were covered by in-house funds of the Department of Diagnostic and Interventional Neuroradiology, School of Medicine, Technical University of Munich, Klinikum rechts der Isar, Germany. The authors declare no conflict of interest. The project is not funded.

Financial disclosure statement

The authors have nothing to disclose.

Data availability statement

The data sets used and/or analysed during the current study are available from the corresponding author on reasonable request.

REFERENCES

  1. Agbaje JO, de Casteele EV, Salem AS, Anumendem D, Lambrichts I, Politis C. Tracking of the inferior alveolar nerve: its implication in surgical planning. Clin Oral Investig. 2017;21:2213–2220. doi: 10.1007/s00784-016-2014-x. [DOI] [PubMed] [Google Scholar]
  2. Ayyalusamy A, Vellaiyan S, Subramanian S, Ilamurugu A, Satpathy S, Nauman M, et al. Auto-segmentation of head and neck organs at risk in radiotherapy and its dependence on anatomic similarity. Radiat Oncol J. 2019;37:134–142. doi: 10.3857/roj.2019.00038. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Barr ML, Haveles CS, Rezzadeh KS, Nolan IT, Castro R, Lee JC, et al. Virtual surgical planning for mandibular reconstruction with the fibula free flap: a systematic review and meta-analysis. Ann Plastic Surg. 2020;84:117–122. doi: 10.1097/SAP.0000000000002006. [DOI] [PubMed] [Google Scholar]
  4. Burian E, Probst FA, Weidlich D, Cornelius CP, Maier L, Robl T, et al. MRI of the inferior alveolar nerve and lingual nerve-anatomical variation and morphometric benchmark values of nerve diameters in healthy subjects. Clin Oral Investig. 2019;24:2625–2634. doi: 10.1007/s00784-019-03120-7. [DOI] [PubMed] [Google Scholar]
  5. Burian E, Sollmann N, Ritschl LM, Palla B, Maier L, Zimmer C, et al. High resolution MRI for quantitative assessment of inferior alveolar nerve impairment in course of mandible fractures: an imaging feasibility study. Sci Rep. 2020;10:11566. doi: 10.1038/s41598-020-68501-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Duttenhoefer F, Mertens ME, Vizkelety J, Gremse F, Stadelmann VA, Sauerbier S. Magnetic resonance imaging in zirconia-based dental implantology. Clin Oral Implants Res. 2015;26:1195–1202. doi: 10.1111/clr.12430. [DOI] [PubMed] [Google Scholar]
  7. Eley KA, McIntyre AG, Watt-Smith SR, Golding SJ. ‘Black bone’ MRI: a partial flip angle technique for radiation reduction in craniofacial imaging. Br J Radiol. 2012;85:272–278. doi: 10.1259/bjr/95110289. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Eley KA, Watt-Smith SR, Golding SJ. Three-dimensional reconstruction of the craniofacial skeleton with gradient echo magnetic resonance imaging (‘black bone’): what is currently possible. J Craniofac Surg. 2017;28:463–467. doi: 10.1097/SCS.0000000000003219. [DOI] [PubMed] [Google Scholar]
  9. Eley KA, Watt-Smith SR, Sheerin F, Golding SJ. ‘Black Bone’ MRI: a potential alternative to CT with three-dimensional reconstruction of the craniofacial skeleton in the diagnosis of craniosynostosis. Eur Radiol. 2014;24:2417–2426. doi: 10.1007/s00330-014-3286-7. [DOI] [PubMed] [Google Scholar]
  10. Flugge T, Hovener JB, Ludwig U, Eisenbeiss AK, Spittau B, Hennig J, et al. Magnetic resonance imaging of intraoral hard and soft tissues using an intraoral coil and FLASH sequences. Eur Radiol. 2016;26:4616–4623. doi: 10.1007/s00330-016-4254-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Gelaude F, Vander Sloten J, Lauwers B. Accuracy assessment of CT-based outer surface femur meshes. Comput Aided Surg. 2008;13:188–199. doi: 10.3109/10929080802195783. [DOI] [PubMed] [Google Scholar]
  12. Gersing AS, Pfeiffer D, Kopp FK, Schwaiger BJ, Knebel C, Haller B, et al. Evaluation of MR-derived CT-like images and simulated radiographs compared to conventional radiography in patients with benign and malignant bone tumors. Eur Radiol. 2018;29:13–21. doi: 10.1007/s00330-018-5450-y. [DOI] [PubMed] [Google Scholar]
  13. Geusens J, Sun Y, Luebbers HT, Bila M, Darche V, Politis C. Accuracy of computer-aided design/computer-aided manufacturing-assisted mandibular reconstruction with a fibula free flap. J Craniofac Surg. 2019;30:2319–2323. doi: 10.1097/SCS.0000000000005704. [DOI] [PubMed] [Google Scholar]
  14. Goto TK, Nishida S, Nakamura Y, Tokumori K, Nakamura Y, Kobayashi K, et al. The accuracy of 3-dimensional magnetic resonance 3D vibe images of the mandible: an in vitro comparison of magnetic resonance imaging and computed tomography. Oral Surg Oral Med Oral Pathol Oral Radiol Endod. 2007;103:550–559. doi: 10.1016/j.tripleo.2006.03.011. [DOI] [PubMed] [Google Scholar]
  15. Haddock NT, Monaco C, Weimer KA, Hirsch DL, Levine JP, Saadeh PB. Increasing bony contact and overlap with computer-designed offset cuts in free fibula mandible reconstruction. J Craniofac Surg. 2012;23:1592–1595. doi: 10.1097/SCS.0b013e318257576c. [DOI] [PubMed] [Google Scholar]
  16. Hanasono MM, Skoracki RJ. Computer-assisted design and rapid prototype modeling in microvascular mandible reconstruction. Laryngoscope. 2013;123:597–604. doi: 10.1002/lary.23717. [DOI] [PubMed] [Google Scholar]
  17. Hilgenfeld T, Kastel T, Heil A, Rammelsberg P, Heiland S, Bendszus M, et al. High-resolution dental magnetic resonance imaging for planning palatal graft surgery-a clinical pilot study. J Clin Periodontol. 2018;45:462–470. doi: 10.1111/jcpe.12870. [DOI] [PubMed] [Google Scholar]
  18. Huotilainen E, Jaanimets R, Valasek J, Marcian P, Salmi M, Tuomi J, et al. Inaccuracies in additive manufactured medical skull models caused by the DICOM to STL conversion process. J Craniomaxillofac Surg. 2014;42:e259–265. doi: 10.1016/j.jcms.2013.10.001. [DOI] [PubMed] [Google Scholar]
  19. Juerchott A, Freudlsperger C, Weber D, Jende JME, Saleem MA, Zingler S, et al. In vivo comparison of MRI- and CBCT-based 3D cephalometric analysis: beginning of a non-ionizing diagnostic era in craniomaxillofacial imaging. Eur Radiol. 2020;30:1488–1497. doi: 10.1007/s00330-019-06540-x. [DOI] [PubMed] [Google Scholar]
  20. Kraeima J, Dorgelo B, Gulbitti HA, Steenbakkers R, Schepman KP, Roodenburg JLN, et al. Multi-modality 3D mandibular resection planning in head and neck cancer using CT and MRI data fusion: a clinical series. Oral Oncol. 2018;81:22–28. doi: 10.1016/j.oraloncology.2018.03.013. [DOI] [PubMed] [Google Scholar]
  21. Mastrogiacomo S, Dou W, Jansen JA, Walboomers XF. Magnetic resonance imaging of hard tissues and hard tissue engineered bio-substitutes. Mol Imaging Biol. 2019;21:1003–1019. doi: 10.1007/s11307-019-01345-2. [DOI] [PubMed] [Google Scholar]
  22. McGrath H, Li P, Dorent R, Bradford R, Saeed S, Bisdas S, et al. Manual segmentation versus semi-automated segmentation for quantifying vestibular schwannoma volume on MRI. Int J Comput Assist Radiol Surg. 2020;15:1445–1455. doi: 10.1007/s11548-020-02222-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Metzler P, Geiger EJ, Alcon A, Ma X, Steinbacher DM. Three-dimensional virtual surgery accuracy for free fibula mandibular reconstruction: planned versus actual results. J Oral Maxillofac Surg. 2014;72:2601–2612. doi: 10.1016/j.joms.2014.07.024. [DOI] [PubMed] [Google Scholar]
  24. Probst FA, Burian E, Malenova Y, Lyutskanova P, Stumbaum MJ, Ritschl LM, et al. Geometric accuracy of magnetic resonance imaging-derived virtual three-dimensional bone surface models of the mandible in comparison to computed tomography and cone beam computed tomography: a porcine cadaver study. Clin Implant Dent Relat Res. 2021;23:779–788. doi: 10.1111/cid.13033. [DOI] [PubMed] [Google Scholar]
  25. Probst FA, Schweiger J, Stumbaum MJ, Karampinos D, Burian E, Probst M. Magnetic resonance imaging based computer-guided dental implant surgery – a clinical pilot study. Clin Implant Dent Relat Res. 2020;22:612–621. doi: 10.1111/cid.12939. [DOI] [PubMed] [Google Scholar]
  26. Probst M, Richter V, Weitz J, Kirschke JS, Ganter C, Troeltzsch M, et al. Magnetic resonance imaging of the inferior alveolar nerve with special regard to metal artifact reduction. J Craniomaxillofac Surg. 2017;45:558–569. doi: 10.1016/j.jcms.2017.01.009. [DOI] [PubMed] [Google Scholar]
  27. Raudaschl PF, Zaffino P, Sharp GC, Spadea MF, Chen A, Dawant BM, et al. Evaluation of segmentation methods on head and neck CT: auto-segmentation challenge 2015. Med Phys. 2017;44:2020–2036. doi: 10.1002/mp.12197. [DOI] [PubMed] [Google Scholar]
  28. Ritschl LM, Kilbertus P, Grill FD, Schwarz M, Weitz J, Nieberler M, et al. In-house, open-source 3D-software-based, CAD/CAM-planned mandibular reconstructions in 20 consecutive free fibula flap cases: an explorative cross-sectional study with three-dimensional performance analysis. Front Oncol. 2021;11:731336. doi: 10.3389/fonc.2021.731336. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Ritschl LM, Roth M, Fichter AM, Mittermeier F, Kuschel B, Wolff KD, et al. The possibilities of a portable low-budget three-dimensional stereophotogrammetry system in neonates: a prospective growth analysis and analysis of accuracy. Head Face Med. 2018;14:11. doi: 10.1186/s13005-018-0168-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Robinson R, Valindria VV, Bai W, Oktay O, Kainz B, Suzuki H, et al. Automated quality control in image segmentation: application to the UK Biobank cardiovascular magnetic resonance imaging study. J Cardiovasc Magn Reson. 2019;21:18. doi: 10.1186/s12968-019-0523-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Rommel N, Kesting MR, Rohleder NH, Bauer FMJ, Wolff K-D, Weitz J. Mandible reconstruction with free fibula flaps: Outcome of a cost-effective individual planning concept compared with virtual surgical planning. J Craniomaxillofac Surg. 2017;45:1246–1250. doi: 10.1016/j.jcms.2017.04.010. [DOI] [PubMed] [Google Scholar]
  32. Saarikko A, Mellanen E, Kuusela L, Leikola J, Karppinen A, Autti T, et al. Comparison of black bone MRI and 3D-CT in the preoperative evaluation of patients with craniosynostosis. J Plast Reconstr Aesthet Surg. 2020;73:723–731. doi: 10.1016/j.bjps.2019.11.006. [DOI] [PubMed] [Google Scholar]
  33. Sakai Y, Okamura K, Kitamoto E, Kami YN, Shirasaka T, Mikayama R, et al. Improved scan method for dental imaging using multidetector computed tomography: a phantom study. Dentomaxillofac Radiol. 2020;49:20190462. doi: 10.1259/dmfr.20190462. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Schepers RH, Raghoebar GM, Vissink A, Stenekes MW, Kraeima J, Roodenburg JL, et al. Accuracy of fibula reconstruction using patient-specific CAD/CAM reconstruction plates and dental implants: a new modality for functional reconstruction of mandibular defects. J Craniomaxillofac Surg. 2015;43:649–657. doi: 10.1016/j.jcms.2015.03.015. [DOI] [PubMed] [Google Scholar]
  35. Siriwanarangsun P, Statum S, Biswas R, Bae WC, Chung CB. Ultrashort time to echo magnetic resonance techniques for the musculoskeletal system. Quant Imaging Med Surg. 2016;6:731–743. doi: 10.21037/qims.2016.12.06. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Smith EJ, Anstey JA, Venne G, Ellis RE. Using additive manufacturing in accuracy evaluation of reconstructions from computed tomography. Proc Inst Mech Eng H. 2013;227:551–559. doi: 10.1177/0954411912474612. [DOI] [PubMed] [Google Scholar]
  37. Sollmann N, Loffler MT, Kronthaler S, Bohm C, Dieckmeyer M, Ruschke S, et al. MRI-based quantitative osteoporosis imaging at the spine and femur. J Magn Reson Imaging. 2021;54:12–35. doi: 10.1002/jmri.27260. [DOI] [PubMed] [Google Scholar]
  38. Stirling Craig E, Yuhasz M, Shah A, Blumberg J, Salomon J, Lowlicht R, et al. Simulated surgery and cutting guides enhance spatial positioning in free fibular mandibular reconstruction. Microsurgery. 2015;35:29–33. doi: 10.1002/micr.22229. [DOI] [PubMed] [Google Scholar]
  39. Succo G, Berrone M, Battiston B, Tos P, Goia F, Appendino P, et al. Step-by-step surgical technique for mandibular reconstruction with fibular free flap: application of digital technology in virtual surgical planning. Eur Arch Otorhinolaryngol. 2015;272:1491–1501. doi: 10.1007/s00405-014-3078-3. [DOI] [PubMed] [Google Scholar]
  40. Suchyta MA, Gibreel W, Hunt CH, Gorny KR, Bernstein MA, Mardini S. Using black bone magnetic resonance imaging in craniofacial virtual surgical planning: a comparative cadaver study. Plast Reconstr Surg. 2018;141:1459–1470. doi: 10.1097/PRS.0000000000004396. [DOI] [PubMed] [Google Scholar]
  41. Suchyta MA, Gibreel W, Sabbagh MD, Hunt CH, Gorny KR, Lu A, et al. Using black bone magnetic resonance imaging for fibula free flap surgical planning: a means to reduce radiation exposure with accurate surgical outcomes. Plast Reconstr Surg 2021;148:77e–82e. [DOI] [PubMed]
  42. Tang NSJ, Ahmadi I, Ramakrishnan A. Virtual surgical planning in fibula free flap head and neck reconstruction: a systematic review and meta-analysis. J Plast Reconstr Aesthet Surg. 2019;72:1465–1477. doi: 10.1016/j.bjps.2019.06.013. [DOI] [PubMed] [Google Scholar]
  43. Tarsitano A, Ciocca L, Scotti R, Marchetti C. Morphological results of customized microvascular mandibular reconstruction: a comparative study. J Craniomaxillofac Surg. 2016;44:697–702. doi: 10.1016/j.jcms.2016.03.007. [DOI] [PubMed] [Google Scholar]
  44. Tong N, Gou S, Yang S, Cao M, Sheng K. Shape constrained fully convolutional DenseNet with adversarial training for multiorgan segmentation on head and neck CT and low-field MR images. Med Phys. 2019;46:2669–2682. doi: 10.1002/mp.13553. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Van den Broeck J, Vereecke E, Wirix-Speetjens R, Vander Sloten J. Segmentation accuracy of long bones. Med Eng Phys. 2014;36:949–953. doi: 10.1016/j.medengphy.2014.03.016. [DOI] [PubMed] [Google Scholar]
  46. van Eijnatten M, van Dijk R, Dobbe J, Streekstra G, Koivisto J, Wolff J. CT image segmentation methods for bone used in medical additive manufacturing. Med Eng Phys. 2018;51:6–16. doi: 10.1016/j.medengphy.2017.10.008. [DOI] [PubMed] [Google Scholar]
  47. van Eijnatten M, van Dijk R, Dobbe J, Streekstra G, Koivisto J, Wolff J. CT image segmentation methods for bone used in medical additive manufacturing. Med Eng Phys. 2018;51:6–16. doi: 10.1016/j.medengphy.2017.10.008. [DOI] [PubMed] [Google Scholar]
  48. Wallner J, Mischak I, Jan E. Computed tomography data collection of the complete human mandible and valid clinical ground truth models. Sci Data. 2019;6:190003. doi: 10.1038/sdata.2019.3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Wallner J, Schwaiger M, Hochegger K, Gsaxner C, Zemann W, Egger J. A review on multiplatform evaluations of semi-automatic open-source based image segmentation for cranio-maxillofacial surgery. Comput Methods Programs Biomed. 2019;182:105102. doi: 10.1016/j.cmpb.2019.105102. [DOI] [PubMed] [Google Scholar]
  50. Yeung AWK, Jacobs R, Bornstein MM. Novel low-dose protocols using cone beam computed tomography in dental medicine: a review focusing on indications, limitations, and future possibilities. Clin Oral Investig. 2019;23:2573–2581. doi: 10.1007/s00784-019-02907-y. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

The data sets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Articles from Oral Health & Preventive Dentistry are provided here courtesy of Quintessence Publishing Co, Inc

RESOURCES