Technical note: Atlas‐based Auto‐segmentation of masticatory muscles for head and neck cancer radiotherapy

Abstract Purpose The study aimed to use quantitative geometric and dosimetric metrics to assess the accuracy of atlas‐based auto‐segmentation of masticatory muscles (MMs) compared to manual drawn contours for head and neck cancer (HNC) radiotherapy (RT). Materials and methods Fifty‐eight patients with HNC treated with RT were analyzed. Paired MMs (masseter, temporalis, and medial and lateral pterygoids) were manually delineated on planning computed tomography (CT) images for all patients. Twenty‐nine patients were used to generate the MM atlas. Using this atlas, automatic segmentation of the MMs was performed for the remaining 29 patients without manual correction. Auto‐segmentation accuracy for MMs was compared using dice similarity coefficients (DSCs), Hausdorff distance (HD), HD95, and variation in the center of mass (∆COM). The dosimetric impact on MMs was calculated (∆dose) using dosimetric parameters (D99%, D95%, D50%, and D1%), and compared with the geometric indices to test correlation. Results DSCmean ranges from 0.79 ± 0.04 to 0.85 ± 0.04, HDmean from 0.43 ± 0.08 to 0.82 ± 0.26 cm, HD95mean from 0.32 ± 0.08 to 0.42 ± 0.16 cm, and ∆COMmean from 0.18 ± 0.11 to 0.33 ± 0.23 cm. The mean MM volume difference was < 15%. The correlation coefficient (r) of geometric and dosimetric indices for the four MMs ranges between −0.456 and 0.300. Conclusions Atlas‐based auto‐segmentation for masticatory muscles provides geometrically accurate contours compared to manual drawn contours. Dose obtained from those auto‐segmented contours is comparable to that from manual drawn contours. Atlas‐based auto‐segmentation strategy for MM in HN radiotherapy is readily availalbe for clinical implementation.

outcomes in the setting of pathogen-associated HNCs, reducing long-term toxicity is increasingly important. Specifically, trismus due to RT-induced masticatory muscles (MMs) injury is a common clinical complication for patients with HNC treated with RT and has a significant impact on health-related quality of life. [2][3][4] However, contouring of these muscles as dose-limiting structures during radiotherapy treatment planning is not routine. Due to low soft-tissue contrast and lack of clear boundaries on typical planning CT images, manual segmentation of these muscles as organs-at-risk (OARs) is challenging. Thus, manual segmentation for these muscles is subject to large interuser variability and is time consuming. 5,6 As a result, RT dose tolerance levels to MMs have not been well studied. To reduce toxicity and improve long-term patient swallowing outcomes and quality of life, dose constraints for the MMs need to be established and accurate, consistent delineation of MMs is necessary.
In recent years, with improved technology for medical image analysis, computer-aided fully or semiautomatic segmentation techniques have shown promise in radiation oncology to provide fast and accurate OAR segmentation. [7][8][9] Atlas-based auto-segmentation is an important, automatic segmentation technique which uses atlas templates built from previously validated OAR contours to automatically create contours for new patients. [9][10][11] The key component of auto-segmentation is a database (i.e., the so-called atlas) containing image data with OAR segmentations. The atlas contours are then propagated to the image data of a new patient via rigid and deformable image registrations. Several studies have evaluated the use of atlas-based auto-segmentation to reduce contouring time and interand intraobserver variations in OARs contouring for HNC. 5,12 A few studies 13-15 evaluated atlas-based algorithms in delineating masticatory muscles for HNC patients. However, these studies evaluated their in-house algorithms, which cannot be directly translated to clinical practice and are not widely available. Furthermore, the dataset sample sizes for the previous studies are relatively low. Teguh et al. 13 used ten cases for building the atlas and 12 for testing. Additionally, the dosimetric impact of using auto-segmented contours has not been thoroughly investigated.
We assessed the feasibility of using a commercial atlas-based algorithm for segmenting MMs for HNC patients in a large patient cohort. We also validated the geometric and dosimetric accuracy of auto-segmented contours against manual segmentations performed by experienced HNC radiation oncologists. The major significance of the present study is to establish efficient and accurate MM autosegmentation strategy in the clinical workflow for improving dose-volume assessment. The proposed approach and results are widely available and deployable, thus enabling for future evaluation and optimization of treated patients' Quality of Life (QoL).

2.A | Patients
This single-center retrospective study was performed following Institutional Research Board (IRB) committee approval. A total of 58 patients who received RT treatment with pathologically confirmed squamous cell HNC without MM invasion were included. Demographic data are shown in Table 1. There were 27 patients aged more than 60 yr, and 44 patients were men. Disease sites include 37 cases of oropharynx cancer, seven of larynx cancer, six of nasopharynx and sinonasal cancer, and eight of cancers of other sites. Twenty-four patients had T3/T4 disease; 37 patients had N + disease. All patients were staged I-IV according to the 8th AJCC staging system. The prescription dose for high-risk regions ranges from 60 Gy to 70 Gy, which was delivered via volumetric modulated arc therapy (VMAT).

2.B | Target delineation
All patients were immobilized in the supine position using a head, neck, and shoulder thermoplastic mask with MoldCare pillow and head holder. Simulation CT images with contrast were obtained prior (masseter, temporalis, and medial and lateral pterygoids) was completed by a radiation oncology attending specialized in HNC according to the institutional guidelines and published recommendations. 16 Segmentations were reviewed, edited when needed, and confirmed by a senior HNC radiation oncology attending.

2.C | Atlas-based algorithms
Patient data were divided into two groups for atlas construction (n = 29) and atlas validation (n = 29). A multi-atlas-based autosegmented algorithm (MABS) in Raystation was used to generate autosegmented contours, and multiple atlas templates were fused for testing image datasets. 17 Raystation deformable image registration (DIR) algorithms were used to fuse and deform the two images and propagates contours from multiple datasets to the testing image for

AS. Raystation's ANAtomically Constrained Deformation Algorithm
(ANACONDA) was used for DIR. ANACONDA combines image intensity information with anatomical information in calculating deformation vectors to achieve best match between images. 18 19 . Quantitative volume comparison was defined as:

2.D | Volume comparison and overlap analysis
where Vmean, MS defines the mean manual segmentation volume, and Vmean, AS defines to the mean auto-segmentation volume.
Dice similarity coefficient (DSC) is a geometric volumetric similarity measure used to determine the degree of overlap of two set of contours, which provides a value that simultaneously quantifies differences in volume size and orientation for nonsymmetric shape of contours. DSC normalizes the intersection volume to a value between 0 (no overlap) and 1 (perfect overlap) and is defined as: where Vm and Va are the volumes of the manual drawn and autosegmented contours, respectively 20 .
The Hausdorff distance is another measure of relative contour overlap and defines as the maximum distance of the same object.
x(1,2), y(1, 2), and z(1,2) were indicate the coordinates of the selected reference contours, which is same as the geometric centroid of the contour.

2.E | Dosimetric comparison
Dose-volume information for MM structures was compared, including D99% (the minimum absorbed dose, cGy), D95% (the prescribed dose, cGy), D50% (the median absorbed dose, cGy), and D1% (the maximum absorbed dose, cGy). Dose difference between manual and auto-segmented contours for MMs was calculated as: with dose MS defines the dose of manual segmentation contours, and dose AS defines the dose of auto-segmentation contours.

2.F | Correlation of volume accuracy with plan quality
To explore the correlation between contour accuracy and plan quality, geometric indices were analyzed with respect to dosimetric endpoints for every MM. The correlation between geometric indices and dose difference was evaluated using Pearson's correlation coefficient (r).

| RESULTS
A representative example of manual and auto-segmented contours is shown in Fig. 1(a). Visually, there is an excellent agreement for each MM. Figure 1(b) shows DVHs for the patient in Fig. 1(a). For the 29 test patients, the geometric comparison indices are shown in Fig. 2 and Table 2.   Table 3 shows that the observed correlation coefficient of geometric and dosimetric indices for MMs. No strong correlation was found amongst geometric indices, while week correlation was observed between some geometric indices and dose differences. The absolute percentage dose differences between manual and auto-segmented contours (difference normalized to the dose value of manual contours) for each MM pair are tabulated in Table 4. The dose-volume parameters evaluated include D99, D95, D50, and D1. The percentage dose differences for all four dose volume parameters are mostly less than 15% for all MM contours, with higher average dose deviations observed with temporalis and lateral pterygoid muscles.
To further elaborate the differences, Fig. 3 shows a linear regression between manual and auto-segmented contours for the four pairs of

| DISCUSSION
In this study, we assessed the performance of atlas-based segmentation for masticatory muscles compared with manual segmentation.  used ten HNC patients for building atlas, and 12 for testing. They showed that DSC between auto-contours and manual contours had a mean of 0.71 for masticatory muscles using a multiple-subject approach. Han et al. 14 tested ten HNC patients by using two atlas selection strategies, and found that the median DSC is below 0.8 for the masseter muscles and pterygoid muscles using a single atlas, but over 0.8 using a multi-atlas strategy. Hague et al. 15

| CONCLUSION
This study is the first to examine the accuracy of atlas-based autosegmentation for MM delineation, using both multiple geometric and dosimetric indices. We found that atlas-based auto-segmentation for muscles of mastication results in geometrically precise automatic organ segmentation and similar organ dose outcomes as compared to manual segmentation. Future work will validate our results on a larger prospective dataset and compare results with other automatic contour generation strategies.

CONF LICT OF I NTEREST
The authors have no relevant conflicts of interest to disclose.