Registration of magnetic resonance and computed tomography images in patients with oral squamous cell carcinoma for three-dimensional virtual planning of mandibular resection and reconstruction.

The aim of this study was to evaluate and present an automated method for registration of magnetic resonance imaging (MRI) and computed tomography (CT) or cone beam CT (CBCT) images of the mandibular region for patients with oral squamous cell carcinoma (OSCC). Registered MRI and (CB)CT could facilitate the three-dimensional virtual planning of surgical guides employed for resection and reconstruction in patients with OSCC with mandibular invasion. MRI and (CB)CT images were collected retrospectively from 19 patients. MRI images were aligned with (CB)CT images employing a rigid registration approach (stage 1), a rigid registration approach using a mandibular mask (stage 2), and two non-rigid registration approaches (stage 3). Registration accuracy was quantified by the mean target registration error (mTRE), calculated over a set of landmarks annotated by two observers. Stage 2 achieved the best registration result, with an mTRE of 2.5±0.7mm, which was comparable to the inter- and intra-observer variabilities of landmark placement in MRI. Stage 2 was significantly better aligned compared to all approaches in stage 3. In conclusion, this study demonstrated that rigid registration with the use of a mask is an appropriate image registration method for aligning MRI and (CB)CT images of the mandibular region in patients with OSCC.

registration approaches (stage 3). Registration accuracy was quantified by the mean target registration error (mTRE), calculated over a set of landmarks annotated by two observers. Stage 2 achieved the best registration result, with an mTRE of 2.5 AE 0.7 mm, which was comparable to the inter-and intra-observer variabilities of landmark placement in MRI. Stage 2 was significantly better aligned compared to all approaches in stage 3. In conclusion, this study demonstrated that rigid registration with the use of a mask is an appropriate image registration method for aligning MRI and (CB)CT images of the mandibular region in patients with OSCC.
Key words: oral squamous cell carcinoma; mandibular reconstruction; X-ray computed tomography; magnetic resonance imaging; multimodal imaging; medical image processing.

Accepted for publication
Oral squamous cell carcinoma (OSCC) is the sixth most common cancer worldwide. The overall 5-year survival rate of patients with OSCC is less than 50% and has not shown any significant improvement over the last decades, despite advances in treatment modalities [1][2][3][4][5] .
Segmental mandibular resection and reconstruction with a free vascularized osseocutaneous flap is currently the recommended treatment for OSCC invading the mandible 6 . The main goal of surgical treatment is to obtain tumour-free resection margins with acceptable remaining function (chewing, swallowing, and speaking) and physical appearance. Achieving tumour-free resection margins is challenging, but crucial for disease control and survival 1,3,[7][8][9][10][11][12] . Recent studies reported that inadequate resection margins were found in 20% of the bone resections, which negatively impacted the 5-year survival rate of patients 1,3,7 . The inability to distinguish tumour from healthy bone tissue intraoperatively during resection is the most common cause of such inadequate margins 10 . Additionally, a re-resection of positive margins in a second operation is not desirable due to technical difficulties, and has a negative effect on the survival of the transplant 11 . During resection, tumour-free surgical margins are the only prognostic factor that the surgeon can control 11 .
The state-of-the-art mandibular reconstruction method is based on preoperative three-dimensional (3D) virtual surgical planning using 3D-printed surgical guides [12][13][14] . With this method, the patient undergoes the necessary imaging, after which the surgeon virtually defines the cutting planes and plans the resection and subsequent reconstruction. Thereafter, the surgical guides are printed and the virtual planning is translated to the surgical procedure. Cutting guides for the mandible and fibula have been shown to provide a better fit of the fibula parts, resulting in a reduced surgical time. However, accurate 3D virtual planning of the surgical cutting guides remains essential in order to achieve a complete resection. Current 3D virtual planning is based on computed tomography (CT) or cone beam CT (CBCT) images, which offer detailed information on bone geometry and cortical bone destruction, but do not provide accurate information on bone marrow involvement and perineural spread of the tumour. In recent years, magnetic resonance imaging (MRI) has increasingly been used for diagnostic purposes due to its better visualization of tumour tissue, mandibular bone marrow involvement, and perineural spread along the inferior alveolar nerve 3,[15][16][17] .
The uncertainty about the location of the tumour boundaries in the 3D virtual planning based on (CB)CT acquisitions could be eliminated by including MRI aquisitions 18 . Overlaying or fusing these MRI and (CB)CT images before the preoperative virtual planning could aid the surgeon in defining the surgical guides and subsequent reconstruction through a more accurate determination of the osteotomy location and a better understanding of the surrounding structures. The integration of fused CT and MRI images for mandibular resection planning in clinical practice has been shown to be a safe and accurate alternative 19 . However, due to the different orientation and position of the mandible in the MRI and (CB)CT acquisitions, an image registration method is required to establish the spatial correspondences between the different images.
The aim of this study was to evaluate and present an automated method to perform image registration of MRI and (CB) CT in the mandibular region in patients with OSCC, which could subsequently be integrated into a pipeline for the virtual planning of mandibular resections and reconstructions.

Dataset
The study was reviewed and approved by the local medical ethics review committee (MEC-2016-143) and was performed in accordance with national and international legislation. The need for informed consent was waived owing to the retrospective and anonymized nature of the study. Preoperative 3D MRI and (CB)CT scans of the head and neck region were collected retrospectively from 19 patients diagnosed between 2014 and 2016 with untreated primary OSCC with invasion of the mandible. The images were anonymized prior to processing. The MRI scans were acquired with a Spin Echo T1-weighted sequence. The in-plane voxel size of MRI was between 0.4 Â 0.4 mm 2 and 0.5 Â 0.5 mm 2 , and slice thickness was between 3 mm and 4 mm. The echo time (TE) ranged from 10.8 ms to 13.6 ms, the repetition time (TR) from 416 ms to 689 ms, and the flip angle (FA) was 90 , 111 , or 160 . The CT imaging in-plane voxel size ranged from 0.3 Â 0.3 mm 2 to 0.5 Â 0.5 mm 2 , and slice thickness from 0.3 mm to 0.6 mm. The CBCT imaging inplane voxel size was 0.3 Â 0.3 mm 2 and slice thickness was 1 mm. The mean time between the MRI and (CB)CT scans was 9 days (range 2-23 days, standard deviation (SD) 5.3 days). No pre-or post-processing was applied to the images and they remained unmodified for the registration.

Registration
An automated image registration method to align the MRI with the (CB)CT images was investigated. The alignment was achieved in three stages. In the first stage, an initial rigid alignment was estimated. Thereafter a more refined rigid alignment was estimated, focused around the mandible. In the third and final stage, a deformable alignment was performed, for which different approaches were compared. In all stages, an automated intensity-based 3D registration framework (Elastix) 20 was used, based on the maximization of mutual information 21 using a stochastic gradient descent optimization method 22 .
In the first stage, two consecutive registrations were performed to achieve an initial rigid alignment. First, a global translation was estimated, since the MRI images needed to be roughly aligned to the (CB)CT image domain. Subsequently, a rigid registration was conducted, estimating both translations and rotations (parameterized by Euler angles).
In the second stage, the initial rigid alignment was fine-tuned by restricting the focus of the algorithm on a 3D region of interest encompassing the mandible, manually drawn in the (CB)CT image. As such, all image information outside of the region of interest was ignored and potential registration difficulties due to pose or appearance changes could be alleviated. The mandibular mask was drawn slice by slice, using open-source ITK-SNAP software 23 .
In the third stage, an evaluation was performed to determine whether the alignment could be refined further using a nonrigid (or deformable) registration to compensate for any geometric distortions in the MRI images 24 . A parametric B-spline freeform deformation model was employed, and (isotropic) control point spacings of 64 mm, 32 mm, and 16 mm were evaluated 25 .
Furthermore, two different approaches for the deformable registration were compared: (a) an asymmetric approach with the (CB)CT image as the fixed (or target, reference) image and the MRI as the moving (or template, source) image; and (b) a symmetric approach in which both images were registered to a common midspace 26 . Results from the literature suggest that symmetric registration techniques can lead to improved registration accuracy and inverse-consistency 27,28 , which are especially critical for treatment planning 29 .
All stages are illustrated in Fig. 1. The three registration stages were executed in a consecutive manner. The registration result after stage 1 was employed to initialize the registration in stage 2. The registration result after stage 2 was employed to initialize the registration in stage 3, where either approach 3a or approach 3b was taken.

Evaluation
Registration accuracy in the mandibular region was evaluated in terms of the mean target registration error (mTRE), by computing the Euclidean distance between corresponding landmarks in MRI and (CB)CT and then averaging it over all landmarks 30 . An extensive landmark set was designed with 39 anatomical reference points in order to evaluate the registration error for the entire mandible specifically. The set consisted of 22 landmarks placed at the roots of each lower tooth (for the molars, both roots were considered as landmarks) and 17 anatomical reference points on the mandible. The set of landmarks is described in detail in Table 1 and Fig. 2. Due to tumour invasion in the bone or removed tooth elements, not all landmarks from the dataset could be annotated for all images.
The landmarks were annotated manually in each MRI and (CB)CT acquisition by two researchers (M.S.H., J.-M.G.) who were trained in advance. The reliability of each landmark was investigated by calculating the inter-observer variability for both MRI and (CB)CT images separately. In addition, one observer (M.S.H.) repeated the annotation of all landmarks, to enable the assessment of the intra-observer variability. Intraand inter-observer variability were quantified by calculating the Euclidean distance (similar to the mTRE) between corresponding landmarks of the same and different observer, respectively.
If the inter-observer variability for a landmark was greater than 5 mm, the landmark was excluded (when applicable both for left and right sides). After the exclusion of the unreliable landmarks, the mTRE was calculated for each registration stage to assess the registration accuracy.

Statistical analysis
Two statistical analyses were performed. First, a comparison was made between the stage that performed best (with the lowest mTRE) and all other registration stages. Second, for the third stage a comparison was made between the asymmetric and symmetric approach for each spacing of the control points (64 mm, 32 mm, and 16 mm). A two-sided Wilcoxon signed rank test at a significance level of 0.05 was used to evaluate the differences in the distribution of the mTRE.

Results
The inter-observer variability for the landmarks gonion (C) and coronoid process (E) was greater than 5 mm (6.5 mm and 6.0 mm, respectively). These landmarks on both sides of the mandible were therefore excluded from the calculation of the mTRE and the observer variabilities. On average, 20 landmarks per patient remained to calculate the mTRE.
The inter-observer variability and intraobserver variability (mean AE SD over all subjects) for landmark placement were found to be 2.4 AE 0.7 mm and 2.0 AE 0.5 mm, respectively, for MRI images and 1.5 AE 0.8 mm and 1.0 AE 0.3 mm, respectively, for (CB)CT images.
After rigid registration in stage 1, the mTRE (mean AE SD over all subjects) was 3.1 AE 1.8 mm. After rigid registration with the use of a mask around the mandible (stage 2), the mTRE was 2.5 AE 0.7 mm. After asymmetric non-rigid registration (stage 3a) with B-spline control point spacings of 64 mm, 32 mm, and 16 mm, the mTREs were 3.6 AE 1.1 mm, 3.4 AE 1.1 mm, and 3.3 AE 1.2 mm, respectively. In the symmetric non-rigid registration (stage 3b), the mTRE values were found to be 3.5 AE 1.2 mm, 3.3 AE 1.2 mm, and 3.1 AE 1.3 mm, respectively. Fig. 3 shows the distributions of mTRE over all subjects for each registration stage, as well as the inter-and intra-observer variability.

Discussion
This study showed that the rigid registration with a mask (stage 2) is the recommended method for registering MRI and (CB)CT images in the mandibular region. Stage 2 achieved a lower mTRE compared to stage 1, although the difference was not significant (P = 0.054). However, the localized focus in stage 2 should generalize better to other patients and be more robust to outliers. A protocol for applying the recommended method is provided in the Appendix A.
In this work, two approaches for deformable registration were applied. A conventional asymmetric approach where the (CB)CT image was employed as the fixed image and the MRI image as the moving image (stage 3a) and a symmetric approach where the images were registered to a common reference system (stage 3b).   Although the differences were not statistically significant, stage 3b achieved a marginal improvement compared to stage 3a for all control point spacings of 64 mm, 32 mm, and 16 mm. Furthermore, the registration error of all approaches in stage 3 was significantly higher than the error in stage 2. As such, the results indicate that non-rigid registration (stage 3) has no added value in this application.
With the large number of landmarks (35 after the exclusions), the mTRE could be estimated reliably. This was even the case for patients for whom not all landmarks could be annotated, e.g. cases where tooth elements were extracted, cases with bone invasion by tumour, and cases with a landmark outside the field of view. The inter-and intra-observer variability of landmark annotations in the (CB)CT images were consistent with those reported in the literature 31,32 . Furthermore, the mTRE achieved by the best registration method (stage 2, 2.5 mm) was similar to the inter-and intra-observer variabilities for MRI images (2.4 mm and 2.0 mm, respectively). As such, lower mTRE values based on the landmarks employed in this study can hardly be expected. Note that a recent study on landmark accuracy in MRI images found lower inter-and intra-observer variability 33 . However, this difference can be explained by the considerably lower slice thickness employed in that previous study (0.53 mm vs at least 3 mm in the present study). Based on the inter-and intra-observer variabilities, the gonion and coronoid process landmarks were excluded; their annotations were most likely hindered by the poor delineation and fuzzy boundaries of the landmark locations 34 . Of note, several other evaluation methodologies for registration accuracy exist, such as the Euclidean distance between centroids, overlap measures, and surface distances of manually segmented anatomical structures 35,36 . In this study, we opted for manual landmark annotations, since a relatively large number of well-defined landmarks could be identified, allowing a reliable estimation of the mTRE, while manual segmentation would have been much more time-consuming.
It appears that data on the registration of MRI and (CB)CT images of the head and neck region are scarce. Previous studies have found registration errors in the range of 1.7-3.3 mm in datasets of four to 16 patients [37][38][39][40][41] . None of these previous studies focused specifically on the mandible, hindering a thorough comparison with the present study. However, the results of these studies suggest that the mTRE of 2.5 mm (for stage 2) achieved indicates a state-of-the-art error level.
Fortunati et al. 42 suggested that patient immobilization during imaging leads to better registration of the MRI and (CB) CT images of the head and neck region. Their study found a registration error of 7.0 mm without immobilization and a registration error of 1.9 mm with immobilization.
The implementation of immobilization equipment for our specific application might not add value in clinical practice, since a competitive mTRE of 2.5 mm was already achieved without immobilization. Moreover, rigorous immobilization of the mandible would be challenging and likely not comfortable for the patient.
The mandibular mask used in stage 2 was drawn manually around the mandible in each slice of the (CB)CT images. Although the mask does not have to be delineated very precisely (it just serves to indicate an approximate region of interest), this manual interaction step may not be desirable in clinical practice. The development of a robust semi-automated or even fully automated segmentation 43 is therefore recommended to accelerate this step. We refer the reader to a recent review of such methodologies for a full overview 44 .
Although the open-source Elastix software was used to implement the image registrations in this study, other (opensource or commercial) software applications that implement similar registration algorithms based on maximization of mutual information could have been used as well. Some well-known open-source examples include NiftyReg 45 and ITKv4/ANTS 46 . After proper configuration, these tools are expected to achieve similar registration accuracy.
Correct determination of the osteotomy location depends on several factors (e.g. the waiting time between imaging and surgery, the process of the translation from 3D virtual planning to the patient, the accurate placement of the cutting guide during surgery, the length of the fibula reconstruction, the relationship of the tumour to the mental nerve and the remaining teeth), which could independently contribute to positive resection margins. For example, several studies have shown that CBCT results in less accurate 3D planning models than CT 47 . When the proposed registration method is translated into clinical practice, image registration errors have to be considered in conjunction with all other sources of errors.
No clinical outcome criteria were employed in this work, such as the resection margin, the frequency of revising the surgical guide during planning, or frequency of local tumour progression, as this was a retrospective study. A randomized clinical trial is needed to determine the added value of registered MRI and CB(CT) in virtual planning of mandibular resection and reconstruction.
This study presented an image registration method for aligning MRI and (CB)CT images of the mandibular region in patients with OSCC. It showed that rigid registration within a region of interest drawn around the mandible is the recommended registration method for the alignment of MRI and (CB)CT images in the mandibular region.
transformation followed by a translation and rotation, to roughly align the anatomical structures in the acquisitions. Store the resulting transformation to initialize stage 2.
3) Perform rigid registration with masked field of view (stage 2). Initialize a new registration with the result from stage 1 and limit the field of view with the binary registration mask. Optimize for the translation and rotation transformation.