Automated atlas-based segmentation for skull base surgical planning

Abstract

Purpose:

Computational surgical planning tools could help develop novel skull base surgical approaches that improve safety and patient outcomes. This defines a need for automated skull base segmentation to improve the usability of surgical planning software. The objective of this work was to design and validate an algorithm for atlas-based automated segmentation of skull base structures in individual image sets for skull base surgical planning.

Methods:

Advanced Normalization Tools software was used to construct a synthetic CT template from 6 subjects, and skull base structures were manually segmented to create a reference atlas. Landmark registration followed by Elastix deformable registration were applied to the template to register it to each of the 30 trusted reference image sets. Dice coefficient, average Hausdorff distance, and clinical usability scoring were used to compare the atlas segmentations to those of the trusted reference image sets.

Results:

The mean for average Hausdorff distance for all structures was less than 2 mm (mean for 95^th percentile Hausdorff distance was less than 5mm). For structures greater than 2.5mL in volume, the average Dice coefficient was 0.73 (range 0.59-0.82), and for structures less than 2.5 mL in volume the Dice coefficient was less than 0.7. The usability scoring survey was completed by three experts, and all structures met the criteria for acceptable effort except for the foramen spinosum, rotundum, and carotid artery, which required more than minor corrections.

Conclusion:

Currently available open-source algorithms, such as the Elastix deformable algorithm, can be used for automated atlas-based segmentation of skull base structures with acceptable clinical accuracy and minimal corrections with the use of the proposed atlas. The first publicly available CT template and anterior skull base segmentation atlas being released (available at this link: http://hdl.handle.net/1773/46259) with this paper will allow for general use of automated atlas-based segmentation of the skull base.

Background

Endoscopic surgery for skull base lesions has been widely adopted over the past two decades. Compared to previously used transcranial approaches to skull base malignances, endoscopic surgical approaches have similar rates of gross total resection and survival. Furthermore, endoscopic approaches provide enhanced visualization of the target and critical structures, reduced complication rates, and shorter hospital stays.[1]

Advances in the field of endoscopic skull base surgery have led to the introduction of a variety of surgical approaches. Over 20 approaches to the skull base, including transorbital[2], transmaxillary[3], transsphenoidal, transethmoidal, transsellar, transclival,[4] or a combination of two or more of these approaches have been well established in the literature.[5] The selection of the optimal surgical approach to a given lesion may be readily apparent for a midline pituitary lesion, but less obvious for lesions located away from the midline. The optimal approach may not be easily determined due to anatomic constraints and variability in lesion configuration and size. Data driven medicine offers an opportunity for complex surgical planning with use of computer-aided algorithms to objectively compare different approaches, which could lead to improved safety and efficiency of surgery.

In an earlier study, a multiobjective cost function was used to guide pre-operative planning for skull base surgery.[6] Key skull base structures were manually segmented on preoperative computed tomography (CT) image sets. Morbidity costs were then assigned to each of the predetermined structures by surgeons, and a weight-based cost function determined a lowest cost surgical approach by numerical optimization. Resultant surgical approach pathways identified by the algorithm were found to be similar to state of the art clinical approaches performed on patients based on expert surgeon review.[6]

However, use of the algorithm in the previous study [6] required significant manual input and precious expert time to produce reliable results. To utilize individual patient image sets in the algorithm, critical anatomic structures were manually segmented and tissue types were assigned to each voxel on each CT by skull base surgeons.[6] Manually segmenting and identifying surgical structures on each patient image sets requires familiarity with 3D software, increases pre-operative planning time, and decreases the generalizability of the algorithm in all clinical settings.

Deformable registration is an existing computational technique to apply information from one image set to another and is particularly effective when there is individual anatomic variability in the size and configuration of structures. By registering of registering a reference atlas to a patient’s CT image sets, anatomic information for every voxel can be transferred from the atlas to the patient’s image set by an automated algorithm. There are many deformable registration algorithms which have been used in different anatomic regions including the brain[7-9],lungs[10,11], breast[12] and for neurosurgical[7] and craniofacial[13] surgical planning. Open-source registration programs include Elastix[14], BRAINSFit[15], and Plastimatch[16] algorithms, which can be used to register a single atlas[17] or multiple atlases[18] to individual patient image sets.

There is a paucity of literature on deformable skull base anatomy registration algorithms for the purpose of surgical planning. The skull base is a difficult region to register due to variability in size and different tissue types of the structures involved. Previous attempts at deformable registration in the head and neck have identified methods for accurate registration of one or two structures but have not identified a comprehensive algorithm. [19]

In this study, we present the first publicly available anatomical skull base atlas with corresponding synthetic CT template and validation that uses of an open-source deformable registration algorithm and can successfully apply data from the anterior skull base atlas and template to individual patient image sets. The objective was to generate accurate segmentations of the skull base with acceptable values for Hausdorff distance, Dice coefficient, and clinical effort needed for corrections.

Methods

Use of high spatial resolution skull base CT image sets without anterior skull base pathology obtained as part of routine clinical care was approved by Seattle Children’s Hospital Institutional Review Board (SCH IRB # STUDY00001830).

An anonymous skull base CT template was constructed with the goal of accounting for anatomic variation and public distribution. We identified high quality skull base CT image sets (spacing 0.4-1.2mm) from six subjects (ages 13-26). These image sets were co-registered to a common coordinate system to in a iterative process to create a synthetic CT template through use of Advanced Normalization Tools (ANTs)[20] software, which is open-source deformable image registration framework that is well studied.[21] Because the individual image sets are combined into one image set without identifiable medical information, public release of the synthetic CT template was also approved by the SCH IRB.

3D Slicer (www.slicer.org), an open-source platform for medical image informatics, image processing, and three-dimensional visualization[22], was used to segment surgical skull base structures on the CT template and designate the voxels corresponding to the anterior skull base to create the anterior skull base atlas. The same structures were also manually segmented on 30 other CT image sets to serve as trusted reference data.

The segmented atlas was then assessed for clinical accuracy. An otolaryngologist and a neuroradiologist conferred until consensus was reached on the segmentation voxel assignment for each of the following structures: ocular globe, bone of anterior skull base, optic nerve, carotid, septum, nasolacrimal duct, extraocular muscles, foramen ovale, foramen spinosum, foramen rotundum, vidian canal, pterygopalatine fossa, pituitary, and cavernous sinus.

Registration

Prior to deformable registration of the atlas to each of the individual trusted reference image sets, pre-processing steps were performed, including removal of background structures such as CT gantry material, nasal cannula, and other non-anatomic features, setting backgrounds thresholds to the same intensity values, and cropping CT image sets to the same anatomic levels (superior limit of frontal sinus to inferior limit of maxilla).

Four landmarks (anterior/superior extent of bilateral zygomaticofrontal sutures, tip of the nasal bones in midline, and posterior edge of the sella in midline) were used as starting points for automatic rigid registration of the CT template to each subject image set. The template was then resampled based on the rigid registration transform derived from this algorithm. In the next step, the open-source deformable registration algorithm Elastix was applied to improve the rigid registration. To improve anterior skull base registration for smaller structures, a second iteration of Elastix registration was performed with anterior head mask. The development of the registration algorithm is depicted as a schematic (Figure 1).

Figure 1.

Schematic of the process for atlas-based segmentation.

Validation

The registration process was tested for clinical accuracy by comparing the automated segmentation from the algorithm to the manual segmentation of each trusted reference image sets. Many measures of validation for medical image segmentation have been extensively discussed in the literature.[23,11]

The Dice coefficient is calculated to compute the degree of volumetric overlap of the assignments by manual segmentation. The range for Dice coefficient value is 0 to 1 [24-26]. Hausdorff distance is the greatest of all distances from a point in one set to the closest point in the other set.[23,27] Average Hausdorff distance is the Hausdorff distance averaged over all points so that is it is less sensitive to outliers and has been used as metric for comparison in previous studies.[28-30] The shorter the average Hausdorff distance, the better the match between two segmentations. 95^th percentile Hausdorff distance values were included to better characterize outliers. The Slicer RT extension in 3D Slicer was used to calculate all of the metrics.[31]

We also included a metric of clinical effort required to use the segmentation. Three otolaryngologists used the rubric in Table 1 (adapted from Schreier et al. [32]) to evaluate the clinical usability of the segmentations. The segmentation was considered acceptable if in 80% of the subject image sets the structure receives a score of 2 or 3, meaning minimal or no effort was required to modify the segmentation. Surgeon notes on the quality of segmentations were also recorded.

Table 1.

Criteria for acceptable clinical effort for segmentations

Score	Definition
0	Not acceptable, manual (re)drawing of the entire structure is required
1	Acceptable, major corrections necessary but with acceptable effort
2	Accepted, only minor corrections required
3	Accepted, no corrections required

Power calculations and statistical analysis

Sample size calculations were done to identify how many subject image sets would be needed to capture mean of 0.7 to 1 for the Dice coefficient and 0.5 mm to 2.0 mm for average Hausdorff distance. Power analysis determined CT image sets from 30 subjects would be needed to detect the desired ratio at 80% power for the Dice coefficient (0.7, compared with null ratio 1.0). Similarly, it was determined that at least 10 CT subject image sets would be needed at 80% power to detect differences smaller than 2 mm.

Results

The CT image sets of 6 subjects with median age of 18 years (range 13-26) and two-thirds of whom were female without anterior skull base pathology were co-registered and averaged into a single synthetic CT template (axial spacing 1.2mm, sagittal spacing 0.5mm, coronal spacing 0.5mm) through use of ANTs software. Critical skull base structures were segmented using 3D Slicer (Table 2). Axial view of the CT template can be seen in Figure 2 and it shows that even small structures such as the optic nerve were averaged with good resolution and without noise. Figure 2 also shows segmentations of individual structures overlaid on the CT and image of the 3D model. Voxel assignments for each segmentation were confirmed and agreed upon by an otolaryngologist(R.B.) and a neuroradiologist (F.P.). This CT atlas and anterior skull base segmentation file is available for download from UW ResearchWorks at this site: http://hdl.handle.net/1773/46259.

Table 2.

Quantitative metrics – Mean and standard deviation

	Average volume (cc)	Average Hausdorff distance (mm)	95th percentile Hausdorff distance (mm)	Dice
		Mean (Std dev)	Mean (Std dev)	Mean (Std dev)
Ocular globes	7.75	1.31 (0.4)	2.92 (0.5)	0.80 (0.1)
Bone	183.30	1.53 (0.7)	4.58 (2.1)	0.62 (0.1)
Optic nerve	0.86	1.12 (0.5)	3.01 (1.7)	0.44 (0.1)
Carotid	0.23	1.34 (0.5)	2.77 (0.8)	0.31 (0.1)
Septum	8.05	1.03 (0.5)	2.72 (0.7)	0.70 (0.1)
Nasolacrimal duct	0.36	1.39 (0.6)	3.67 (1.2)	0.34 (0.1)
Extraocular muscles	4.05	1.41 (0.3)	5.08 (1.1)	0.43 (0.1)
Foramen ovale	0.12	0.95 (0.3)	2.22 (0.5)	0.39 (0.1)
Foramen spinosum	0.02	1.32 (0.5)	2.41 (0.7)	0.11 (0.1)
Foramen rotundum	0.04	1.19 (0.6)	2.44 (0.9)	0.18 (0.1)
Vidian canal	0.07	1.39 (0.6)	2.90 (0.9)	0.21 (0.1)
Pterygopalatine fossa	0.73	1.22 (0.5)	3.30 (1.1)	0.47 (0.1)
Pituitary	0.52	1.07 (0.4)	2.47 (0.6)	0.51 (0.1)
Cavernous sinus	0.68	1.04 (0.3)	2.58 (0.6)	0.42 (0.1)

Figure 2.

Left: Axial view of CT template. Second from the left: Segmentation atlas overlaid over CT template. Right two panels: 3D views of segmentation atlas.

The CT template was then deformably registered to each of the 30 trusted reference image sets for validation of the skull base registration algorithm. Each anatomical structure was compared between the atlas and each trusted reference image set by calculating average Hausdorff distance, 95^th percentile Hausdorff distance, and Dice coefficient. The results can be seen in Figure 3 with box plots identifying median, interquartile range, and extremes for each structure.

Figure 3.

Comparison of segmentations after registration of CT template to trusted reference image sets with (a) average Hausdorff distance, (b) 95^th percentile Hausdorff distance, and (c) Dice coefficient. NLD = nasolacrimal duct, EOM = extraocular muscles, f.o. = foramen ovale, f.s. = foramen spinosum, f.r. = foramen rotundum, PPF = pterygopalatine fossa, c.s. = cavernous sinus

Means with standard deviations for average Hausdorff distance, 95^th percentile Hausdorff distance, and Dice coefficient were calculated (Table 2). The mean for average Hausdorff distance for all structures was less than 2 mm, even for smaller structures such as the optic nerve and foramen ovale. The mean for the Dice coefficient was greater than or equal to 0.7 for ocular globes and septum. 95^th percentile and maximum Hausdorff distance were noted to have a larger range particularly in larger structures such as bone and extraocular muscles (Table 2 and Supplement).

Directing attention to the smaller skull base structures (optic nerve, nasolacrimal duct, extraocular muscles, foramen ovale, foramen spinosum, foramen rotundum, vidian canal, pterygopalatine fossa, pituitary, cavernous sinus, and carotid artery), three otolaryngologists clinically validated the atlas registrations for the trusted reference image sets by using the criteria defined in Table 1. As shown in Figure 4, eight out of eleven of these structures were segmented accurately within clinically acceptable effort (the structure receives a score of 2 or 3 for ≥ 80% of the patients). However, foramen spinosum, rotundum, and carotid artery structures scored a 2 or 3 in 70% of cases, requiring more than minor corrections in 30% of cases.

Figure 4.

Acceptability of clinical effort required for clinical use the segmentation (percentage of segmentations that scored 2 or 3 for each structure)

Discussion

There are many commercial and open-source head and neck atlases available for educational purposes, including the Nowinski atlas[33], VOXEL-MAN[34], and the MIDA model[35]. As useful as these atlases are for learning three-dimensional skull base anatomy, none have been published with corresponding imaging data, which prevents their use for image registration and other clinical applications.

To assess feasibility of automated segmentation of the skull base and allow for public release of the CT template, we co-registered the CT image sets of 6 different subjects to create a synthetic template using the ANTs algorithm. A synthetic CT template has been shown to be superior to a single CT atlas in head and neck image registration.[36,37] One study by Sjöberg et al. showed that the time needed to correct automatic segmentation of head and neck lymph nodes was reduced when using a fused atlas compared to single atlas.[37] CTs from 6 subjects were chosen based on a preliminary query of the skull base database. The process was repeated for 12 subjects and the performance was similar, so the template and atlas selected for release comprised of CT data from 6 subjects. A synthetic CT template also allows us to publicly release the CT template DICOM data with the segmentation atlas as the subjects who contributed to the image set cannot be identified. This release of first publicly available skull base atlas with imaging was approved by the SCH IRB. Therefore, the atlas can be modified by end users and other investigators as needed and used for numerous other applications.

Prior to our study, deformable registration algorithms had not been clinically validated for the anterior skull base. Although groups have attempted to identify deformable registration algorithms for the head and neck[17,19,36], none have explored the skull base as its own entity with the goal of surgical planning. With this project, we validated an open-source atlas-based registration algorithm, Elastix, for automatic segmentation of skull base structures. Elastix is based on the Insight Segmentation and Registration Toolkit (ITK) [14,38]. Our clinical validation goals were to identify an algorithm for which all structures were segmented with acceptable values for Hausdorff distance, Dice coefficient, and clinical effort needed for corrections. The technique utilizes rigid landmark registration followed by Elastix registration for segmentation. The mean for average Hausdorff distance for all structures was less than 2 mm (mean for 95^th percentile Hausdorff distance was less than 5mm). This level of accuracy for very small structures such as the skull base foramina transmitting cranial nerves is very useful as the conventional identification of these structures on imaging can be difficult secondary to anatomic variation.

There were two instances in which the manual segmentation of the foramen ovale and spinosum were able to be improved by the automated segmentation. When the average Hausdorff distance was noted to be high, the manual segmentations were reassessed, and it was noted that the registered segmentation had identified aspects of the foramens that had not been completely segmented in the manual segmentation. This highlights a potential role for the atlas and skull base registration to improve manual segmentation, a valuable tool in research, educational, and, most critically, clinical scenarios.

The Dice similarity coefficient is the most widely used metric in validating medical volume segmentations.[23] The Dice coefficient is known to have limited utility for structures that are small, and this was consistent in our study. In such small structures, even a slight discrepancy in location can significantly decrease volumetric overlap. As the Dice coefficient is known to have limitations for smaller structures, the average Hausdorff distance was used as another quantitative metric and was shown to be less 2mm for all structures. As expected the maximum and 95^th percentile Hausdorff distances for the structures had larger values than average Hausdorff distances as they include outliers and most structures are bilateral increasing the maximum distance.

Although these technical metrics are valuable to compare different methods of registration, they do not necessary represent the utility of the registration algorithm for clinical applications. Despite the range, outliers, and low Dice coefficient for some of the structures, surgeons still felt that minimal effort was needed to correct those segmentations to be used clinically. For our purposes, the immediate clinical application is creating a detailed plan for surgery. Indeed, this semi-automated deformable registration algorithm enables a quantitative and robust analysis of anatomy to plan a surgical approach while minimizing risk to critical structures. In sinus and skull base surgery, image navigation permits visualization of instruments within the preoperative CT. The accuracy of this technique is 2 mm or less.[39] As such, an accuracy of 2 mm is a surgically relevant accuracy threshold. An even more pragmatic metric, is how often a surgeon or neuroradiologist needs to enter manual modifications to the automated segmentation. We computed clinical usability scoring described in a study by Schreier et al.[32] This metric allows surgeons to determine if the manual correction of the automated segmentation results involves acceptable effort or is too burdensome. Our results showed that most (8 out of 11) of the smaller structures with lower than standard Dice coefficients had clinical acceptability and required minimal corrections. Surgeons also noted that the segmentations for structures that did not meet criteria for clinical acceptability, which were the foramen spinosum, foramen rotundum, and carotid artery, were still useful in that they identified the location of the structures within 2mm so that corrections could be made efficiently.

Limitations to use of atlas-based segmentation for the skull base will likely be similar to limitations of other atlas-based segmentation registration algorithms.[40] In patients with extreme anatomical variation due to variation in the development of the paranasal sinuses, clinical accuracy of some structures may be lower and require increased manual correction. The CT template was based on patients ages 13-26 years old and applied to trusted references ages 16-20 years old. Since cranial growth is complete by ages 14-15 years old[41], we feel this suggests the applicability of the atlas to older individuals. However, further studies with older individuals may be needed to support this claim. Further larger cohort studies are also needed to determine the clinical accuracy in patients with anterior skull base pathology, which could also be made more efficient by utilizing deep-learning registration methods[42], which have been shown to have good success in the head and neck region in previous studies[43]. Other methods of segmentation such as multi atlas based segmentation[18,44] could be compared for even further improvements in accuracy.

The applications for automated atlas-based skull base segmentation are numerous. To start, it can be used in conjunction with the multiobjective cost function[6] or other software to enhance surgical planning, or to create modifiable educational resources, or even to generate simulation environments for individual surgeries[45].

Conclusion

We developed a novel synthetic CT template from 6 different CT image sets and a corresponding anterior skull base segmentation atlas. We validated the use of open-source algorithms, specifically landmark registration followed by Elastix deformable registration, for automated atlas-based segmentation of skull base structures with acceptable values for Hausdorff and Dice coefficient. Further, in a pragmatic clinical usability test, minimal effort was required by the clinician to use the segmentation. The atlas and CT template released with this paper will allow for general use of automated atlas-based segmentation. The ability to automatically perform highly accurate skull base registration will enable large-scale outcome studies and facilitate automated surgical planning to guide the surgeon.