Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2022 Oct 4.
Published in final edited form as: Otolaryngol Head Neck Surg. 2022 Feb 8;167(4):731–738. doi: 10.1177/01945998221076801

Automated Extraction of Anatomical Measurements From Temporal Bone CT Imaging

Andy S Ding 1,2, Alexander Lu 1,2, Zhaoshuo Li 3, Deepa Galaiya 1, Masaru Ishii 1, Jeffrey H Siewerdsen 2,3, Russell H Taylor 3, Francis X Creighton 1
PMCID: PMC9357851  NIHMSID: NIHMS1798226  PMID: 35133916

Abstract

Objective.

Proposed methods of minimally invasive and robot-assisted procedures within the temporal bone require measurements of surgically relevant distances and angles, which often require time-consuming manual segmentation of preoperative imaging. This study aims to describe an automatic segmentation and measurement extraction pipeline of temporal bone cone-beam computed tomography (CT) scans.

Study Design.

Descriptive study of temporal bone measurements.

Setting.

Academic institution.

Methods.

A propagation template composed of 16 temporal bone CT scans was formed with relevant anatomical structures and landmarks manually segmented. Next, 52 temporal bone CT scans were autonomously segmented using deformable registration techniques from the Advanced Normalization Tools Python package. Anatomical measurements were extracted via in-house Python algorithms. Extracted measurements were compared to ground truth values from manual segmentations.

Results.

Paired t test analyses showed no statistical difference between measurements using this pipeline and ground truth measurements from manually segmented images. Mean (SD) malleus manubrium length was 4.39 (0.34) mm. Mean (SD) incus short and long processes were 2.91 (0.18) mm and 3.53 (0.38) mm, respectively. The mean (SD) maximal diameter of the incus long process was 0.74 (0.17) mm. The first and second facial nerve genus had mean (SD) angles of 68.6 (6.7) degrees and 111.1 (5.3) degrees, respectively. The facial recess had a mean (SD) span of 3.21 (0.46) mm. Mean (SD) minimum distance between the external auditory canal and tegmen was 3.79 (1.05) mm.

Conclusions.

This is the first study to automatically extract relevant temporal bone anatomical measurements from CT scans using segmentation propagation. Measurements from these models can streamline preoperative planning, improve future segmentation techniques, and help develop future image-guided or robot-assisted systems for temporal bone procedures.

Keywords: temporal bone, automated segmentation, atlas, data set curation

The temporal bone houses a complex geometry of nerves, arteries, veins, and the organs for both hearing and balance that are often within millimeters of each other. Due to the close proximity of critical anatomical structures within the surgical field, preoperative planning strategies are integral in optimizing drilling paths and surgical access. However, variability in orientation, position, and size of relevant anatomy can significantly affect preoperative planning and surgical approach. Currently, anatomical measurements within the temporal bone are typically performed on 2-dimensional (2D) preoperative computed tomography (CT) slices, which inherently lack a comprehensive representation of 3-dimensional (3D) structures. To this end, imaging software is often used to create 3D reconstructions of relevant structures, but this process can be time-consuming and tedious. While surgeons frequently manage straightforward measurements of structures by mentally re-creating a 3D framework of the surgical space from 2D images, more nuanced measurements (eg, size of superior canal dehiscence, distance from cochlea to ampulla in middle fossa approaches, distance and angles between the fundus and posterior canal for retrosigmoid approach) are more difficult to obtain accurately and efficiently. Furthermore, there is increasing research on robotic surgical approaches to the temporal bone, and any future technology in this field requires accurate mathematical representations of patient-specific 3D anatomy. Current methods for achieving this require manual segmentation of multiple structures for reliable trajectory mapping, which is both time-intensive and a rate-limiting step for any future robotic surgical workflows. This study presents and validates an automated method for extraction of clinically significant anatomical measurements from temporal bone CT scans. This method builds on our group’s previously validated automated segmentation propagation pipeline1,2 and has applications for significant preoperative planning and analysis of anatomical relationships within the temporal bone.

Methods

This study was approved by the Johns Hopkins Medicine Institutional Review Board. Deidentified and cropped cone-beam CT scans of patient temporal bones were obtained from the Johns Hopkins picture archiving and communication system in the Department of Otolaryngology–Head and Neck Surgery. Scans with anatomy-altering pathology (eg, cholesteatoma, congenital deformities, trauma) or surgical history in the temporal bone were excluded from this study. The resolution of the scans used in this study was 0.1 mm per voxel length, with image dimensions of 512 × 512 × N voxels, where N refers to the number of axial CT slices for a given image. To build an automated pipeline for anatomical measurement extraction, we extended a previously described registration-based segmentation technique2 to not only segment relevant anatomy but also generate annotations within temporal bone CT images.1 To do so, we created 1 segmentation file with labeled anatomy corresponding to a template temporal bone CT image, as well as a separate annotation file with labels denoting the location of clinically useful anatomical landmarks (eg, umbo of the malleus, short and long processes of the incus, first and second genus of the facial nerve). Using this registration-based segmentation technique, we propagated both segmentation and annotation files to correspond to labeled anatomy and landmarks of new temporal bone CT images.

Segmentation Propagation and Landmark Annotation of the Temporal Bone

An automated segmentation propagation technique was implemented, as previously described,2 to label temporal bone anatomy. A crucial part of this automated pipeline is the creation of a labeled temporal bone atlas. To label a target temporal bone CT, this labeled atlas is deformed using image registration techniques to match the target, effectively mapping its labels onto the target. Instead of using a single-patient temporal bone CT image as the atlas, we generated an “average template” temporal bone that represented the average temporal bone shape among a set of 16 temporal bone CT images. The purpose of generating the average template was to create an atlas that was to minimize individual variation that would be present if using a single-patient temporal bone as the atlas. This average template was created by coregistering 16 temporal bone CTs that were used solely for this purpose and were not included in any other part of this study. Anatomical structures (eg, ossicles, bony labyrinth, facial nerve, and chorda tympani) in the average template were then segmented, along with relevant anatomical landmarks (Table 1). Using the Advanced Normalization Tools (ANTs) library,3 segments and annotations of this average template were then mapped onto an unlabeled target data set with deformable image registration methods to predict anatomical labels for the target image. To minimize bias, a separate set of 52 temporal bone data sets (35 for descriptive statistics of anatomical measurements; 17 for comparisons against ground truth manual segmentations and measurements), which were not included in generating the average template bone template, were designated as target images and appropriately labeled using this segmentation propagation technique.

Table 1.

Relevant Anatomical Structures and Landmarks in Each Manually Labeled Temporal Bone.

ID Structure ID Annotation
1 Bone 1 Malleus: lateral process tip
2 Malleus 2 Malleus: manubrium tip
3 Incus 3 Incus: short process tip
4 Stapes 4 Incus: long process tip
5 Bony labyrinth 5 Incus: long process max diameter
6 IAC 6 Facial nerve: labyrinthine segment
7 Superior vestibular nerve 7 Facial nerve: start of labyrinthine segment
8 Inferior vestibular nerve 8 Facial nerve: first genu
9 Cochlear nerve 9 Facial nerve: second genu
10 Facial nerve 10 Facial nerve: end of mastoid segment
11 Chorda tympani 11 Facial nerve: mastoid segment
12 ICA 12 Chorda tympani: distal posterior segment
13 Sigmoid sinus + dura 13 Chorda tympani: intersection with facial nerve
14 Vestibular aqueduct 14 EAC: superior border
15 Mandible
16 EAC
Open in a new tab

Abbreviations: EAC, external auditory canal; IAC, internal auditory canal; ICA, internal carotid artery.

Anatomical Measurement Extraction From Labeled Landmarks

Out of the 52 temporal bone CTs included in this study, 35 CTs were designated for descriptive statistics to represent typical measurements among normal temporal bones. In-house software was developed to extract relevant anatomical measurements from segmented anatomy and annotated landmarks. For each target image, 3D meshes of anatomical structures and landmarks were generated using the marching cubes method (Figure 1).4 Distance, volume, and angle measurements were then extracted using standard Python packages. For the malleus, the distance between the tip of the lateral process and the umbo of the manubrium was calculated to determine the length of the manubrium (Figure 2A). The length of both the short and long processes of the incus was also determined by calculating the distance between the tips of both processes to the centroid of the incus (Figure 2B). In addition, the distal portion of the incus long process was annotated, with the maximal diameter of this portion calculated using the Vascular Modeling Toolkit extension for 3D Slicer.5 For the facial nerve, the apex (turning point) positions of the first and second genus were annotated, along with the beginning of the labyrinthine segment and end of the mastoid segment, to determine the first and second genu angles (Figure 2C). The angle of the first genu was therefore defined by the proximal point of the labyrinthine segment, apex of the first genu, and apex of the second genu, while the angle of the second genu was defined by the apex of the first genu, apex of the second genu, and distal point of the mastoid segment. The facial recess was determined by using 3 points: (1) the intersection of the chorda tympani and the facial nerve at a standard view for cochlear implantation (ie, allowing for visualization of the round window), (2) the distal point of the posterior canalicular chorda tympani prior to entering the middle ear, and (3) the point on the mastoid segment of the facial nerve closest to the distal posterior canalicular chorda tympani. Using these 3 points, the span of the facial recess was defined as the distance between points 2 and 3, while the angle of the facial recess was defined as the angle between these 3 points using point 1 as the apex. Finally, since these 3 points form a triangle, we used (1) the facial recess angle, (2) the distance between point 1 and point 2, and (3) the distance between point 1 and point 3 to calculate the area of this triangle with the trigonometric law of sines as an approximation for the area of the facial recess (Figure 2D). Two additional distances were calculated among this cohort: the closest distance between the labyrinthine segment of the facial nerve and the cochlea (Figure 2E) and the closest distance between the superior portion of the external auditory canal (EAC) and the tegmen tympani (Figure 2F).

Figure 1.

Figure 1.

Open in a new tab

Three-dimensional meshes of example segmentations and landmarks (denoted in black). (A) Facial nerve landmarks. Surrounding structures shown for context. (B) Ossicle landmarks. (C) Superior portion of the external auditory canal (EAC).

Figure 2.

Figure 2.

Open in a new tab

(A-D) Anatomical measurements of the malleus, incus, facial nerve, and facial recess overlaid on respective three-dimensional meshes. (E, F) Distance between the labyrinthine facial nerve and cochlea, as well as between the external auditory canal (EAC) and tegmen. IAC, internal auditory canal.

Accuracy Analysis of Calculated Anatomical Measurements

To evaluate the accuracy of predicted anatomical measurements, 17 of the 52 temporal bone CTs were designated for manual segmentation, which were conducted by 2 medical trainees with experience in temporal bone anatomy and were verified by the senior author. Landmarks were also annotated by a single reviewer to remove interannotator variability. Anatomical measurements extracted from these manually labeled structures and landmarks were then defined as ground truth measurements. Anatomical measurements for these 17 scans were then extracted using our automated pipeline and compared against ground truth measurements. Statistical analyses were conducted using GraphPad Prism 9.2.0 (Graph-Pad Software).

Results

Descriptive Statistics of Extracted Anatomical Measurements

A total of 35 temporal bones were included for descriptive statistics after automatic segmentation and annotation with our segmentation propagation method. Among this cohort, the mean (SD) malleus manubrium length was 4.39 (0.34) mm. In the incus, mean (SD) lengths for short and long processes were 2.91 (0.18) mm and 3.53 (0.38) mm, respectively, using the centroid of the incus as a reference point. The mean (SD) angle between the short and long process was found to be 102.3 (6.4) degrees. Furthermore, at the typical site of stapedial prosthesis attachment on the incus long process, the mean (SD) maximal diameter was 0.74 (0.17) mm. Landmarks on the facial nerve and chorda tympani were used to calculate measurements relevant to the facial nerve and facial recess. In this cohort of temporal bones, the mean (SD) angle of the first genu of the facial nerve was 68.6 (6.7) degrees. In addition, the mean (SD) angle of the second genu was determined to be 111.1 (5.3) degrees. Using the most distal point on the posterior canalicular segment of the chorda tympani, as well as its corresponding closest point on the mastoid segment of the facial nerve, the mean (SD) span of the facial recess was 3.21 (0.46) mm, with a mean (SD) angle of 27.4 (5.1) degrees. By approximating the shape of the facial recess as a triangle, the mean (SD) area of the facial recess was calculated to be 9.08 (2.34) mm2. Additional interstructure distances were extracted. The mean (SD) closest distance between the labyrinthine segment of the facial nerve and the cochlea was 0.40 (0.05) mm, while the mean (SD) closest distance between the superior portion of the EAC and the tegmen tympani was 3.79 (1.05) mm.

Accuracy Analysis of Anatomical Measurement Extraction With Segmentation Propagation

Anatomical measurements were extracted from a cohort of 17 manually labeled temporal bones using our automated segmentation pipeline. Importantly, unpaired t tests showed no significant difference between predicted measurements from this cohort and predicted measurements from the 35 temporal bones previously described. Multiple paired t tests between predicted and ground truth measurements from this cohort also showed no significant difference, although difference between predicted and ground truth superior EAC-to-tegmen tympani distances trended toward significance (predicted: 3.30 mm vs ground truth: 2.84 mm; difference: 0.46; P = 0.055) (Table 2).

Table 2.

Predicted Anatomical Measurements Compared Against Ground Truth Measurements.

Anatomical measurement Predicted Expected Mean (SD) error, % P value
Malleus
 Manubrium length, mm 4.26 4.43 −0.18 (−3.95) 0.279
Incus
 Short process length, mm 2.98 3.04 −0.06 (−2.06) 0.525
 Long process length, mm 3.39 3.54 −0.14 (−4.05) 0.139
 Interprocess angle, deg 101.50 97.62 3.87 (3.97) 0.279
 Long process max diameter, mm 0.67 0.63 0.04 (5.96) 0.548
Facial nerve, deg
 First genu angle 70.50 64.02 6.48 (10.13) 0.356
 Second genu angle 112.40 113.30 −0.89 (−0.79) 0.553
Facial recess
 Angle, deg 25.43 28.92 −3.49 (−12.07) 0.328
 Area, mm2 9.54 8.91 0.63 (7.08) 0.613
 Span, mm 3.22 3.07 0.15 (4.92) 0.553
Superior EAC-tegmen
 Closest distance, mm 3.30 2.84 0.46 (16.15) 0.055
Facial nerve–cochlea
 Closest distance, mm 0.39 0.36 0.03 (9.14) 0.553
Open in a new tab

Abbreviation: EAC, external auditory canal.

Discussion

This is the first study, to our knowledge, to validate a method for automated anatomical measurement extraction with reliable accuracy. Prior methods for calculating clinically relevant anatomical measurements typically involve manual segmentation of structures on CT imaging with placement of fiducials on visualization software6-8 or meticulous harvesting of cadaveric structures with direct manual measurements.9-11 By manually segmenting a single average bone template along with clinical relevant landmarks, this image registration-based method is able to automatically propagate template labels to other temporal bone scans with reliable accuracy. We believe the ability to automate this process will have applications in current preoperative planning workflows and large-scale studies of temporal bone anatomy, as well as for future work in integrating image-guided robotics into neurotologic surgery. Regarding its usefulness in current preoperative planning, this study included measurements that are useful for a variety of procedures, such as cochlear implantation (facial recess span and angle), stapedectomy (incus long process diameter for prosthesis placement), mastoidectomy (EAC-to-tegmen distance to approximate skull base thickness), and the middle fossa approach for facial nerve decompression (distance from the labyrinthine segment of the facial nerve to the cochlea). This technology not only has the potential to improve existing surgical planning pipelines but can also guide robotic surgical systems in this space by providing detailed anatomical information to both the surgeon and the robot. Taking minimally invasive robotic cochlear implantation as an example, this software can help in both determining the feasibility of a drilling path given the span of the patient’s facial recess and finding the optimal drilling path for a particular patient. The use of this pipeline can further be used for the development of large population data sets to better our understanding of temporal bone anatomy. Creation of these data sets is also helpful in surgical training, simulation development, and design of surgical tools for lateral skull base procedures.

The accuracy of this pipeline is further corroborated by comparing its findings to published findings of manual measurements of similar metrics. Reported anatomical measurements for malleus manubrium length,6,11-13 incus interprocess angle,13 facial nerve first and second genu angles,14-16 facial - nerve–cochlea distance,17 and superior EAC-tegmen distance18 in this study have been consistent with measurements reported in previous literature. However, our methodology for calculating incus short and long process lengths varies from previous definitions. While other studies have defined short and process lengths as the distance between the tip of each process to the head of the incus,6,12,13 we instead calculated the centroid of the incus as a common endpoint. This is because the centroid of the incus can be objectively calculated from its 3D mesh, rather than subjectively identifying a point at the head of the incus. As a result, our reported short and long process lengths for the incus cannot be reliably compared to previous measurements.

While this automated pipeline has consistently demonstrated reliable and accurate extraction of anatomical measurements, there are some limitations with this study. First, this algorithm depends on an adequate amount of input CTs to generate an average temporal bone template. However, this method does not require manual segmentations of these CTs, which are required by deep learning or statistical shape model methods19-22 and only require segmentation of the average temporal bone template. Second, this method currently approximates the facial recess as a rigid triangle and does not account for the curvature of the chorda tympani, leading to less accurate reporting of the angle and area of the facial recess. Nonlinear analyses of the chorda tympani would be useful in more accurately calculating facial recess metrics. Finally, this algorithm uses a template image of slice thickness 0.1 mm, whereas the standard thickness of most routine temporal bone CT scans is about 0.6 mm.23 Although methods exist within the ANTs Python Library to upsample lowerresolution CT scans and therefore preserve the quality of the segmentations produced with a higher-resolution template image,3 we have not verified the accuracy of this pipeline using a more standard resolution template image and plan to in future studies. Despite these limitations, we provide promising applications of this pipeline that result in anatomical measurements with relatively small error when compared to ground truth values.

Conclusion

Current methods for obtaining clinically relevant anatomical measurements require time-intensive manual segmentation of relevant structures or direct measurements of preoperative CT slices that suffer from unreliable accuracy. This study extended a previously validated segmentation pipeline by annotating landmarks in temporal bone anatomy and using those landmarks to extract anatomical measurements in an automated fashion. We have shown that this pipeline performs with reliable accuracy when compared with corresponding ground truth measurements. Measurements extracted from this pipeline can be used to inform surgeons during preoperative planning and to further investigate the clinical significance of these values with respect to patient characteristics. We plan to recruit additional expert labelers to investigate the multiple manual segmentations against the ground truth segmentations used in this study. In doing so, we plan to evaluate whether measurements produced by this pipeline are within interlabeler variability. Furthermore, future work will evaluate differences in anatomical landmarks and measurements stratified by patient age, sex, ethnicity, and symptomatology.

Funding source:

Funding and equipment support was provided by a contract between Galen Robotics and the Johns Hopkins University.

Footnotes

Competing interests: Under a license agreement between Galen Robotics, Inc and the Johns Hopkins University, Russell H. Taylor and the university are entitled to royalty distributions on technology related to technology described in the study discussed in this publication. Russell H. Taylor also is a paid consultant to and owns equity in Galen Robotics, Inc. This arrangement has been reviewed and approved by the Johns Hopkins University in accordance with its conflict-of-interest policies.

This article was presented at the 125th AAO-HNSF 2021 Annual Meeting & OTO Experience; October 5, 2021; Los Angeles, California.

References

  • 1.Ding AS, Lu A, Li Z, et al. Automated registration-based temporal bone computed tomography segmentation for applications in neurotologic surgery [published online September 7, 2021]. Otolaryngol Head Neck Surg. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Sinha A, Leonard S, Reiter A, Ishii M, Taylor RH, Hager GD. Automatic segmentation and statistical shape modeling of the paranasal sinuses to estimate natural variations. Proc SPIE Int Soc Opt Eng. 2016;9784:97840D. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Avants BB, Tustison NJ, Song G, Cook PA, Klein A, Gee JC. A reproducible evaluation of ANTs similarity metric performance in brain image registration. NeuroImage. 2011;54(3):2033–2044. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Lorensen WE, Cline HE. Marching cubes: a high resolution 3D surface construction algorithm. SIGGRAPH Comput Graph. 1987;21(4):163–169. [Google Scholar]
  • 5.Antiga L, Piccinelli M, Botti L, Ene-Iordache B, Remuzzi A, Steinman DA. An image-based modeling framework for patient-specific computational hemodynamics. Med Biol Eng Comput. 2008;46(11):1097–1112. [DOI] [PubMed] [Google Scholar]
  • 6.Bartling ML, Rohani SA, Ladak HM, Agrawal SK. Micro-CT of the human ossicular chain: statistical shape modeling and implications for otologic surgery. J Anat. 2021;239(4):771–781. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Onbas O, Kantarci M, Murat Karasen R, et al. Angulation, length, and morphology of the styloid process of the temporal bone analyzed by multidetector computed tomography. Acta Radiol. 2005;46(8):881–886. [DOI] [PubMed] [Google Scholar]
  • 8.Van Osch K, Allen D, Gare B, Hudson TJ, Ladak H, Agrawal SK. Morphological analysis of sigmoid sinus anatomy: clinical applications to neurotological surgery. J Otolaryngol Head Neck Surg. 2019;48(1):2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Saha R, Srimani P, Mazumdar A, Mazumdar S. Morphological variations of middle ear ossicles and its clinical implications. J Clin Diagn Res. 2017;11(1):AC01–AC04. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Singh A, Irugu DVK, Kumar R, Verma H, Gupta A, Kumari A. Study of morphologic variability of incudostapedial angle and its relation with temporal bone pneumatization. J Otol. 2020;15(2):50–53. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Todd NW, Creighton FX. Malleus and incus: correlates of size. Ann Otol Rhinol Laryngol. 2013;122(1):60–65. [DOI] [PubMed] [Google Scholar]
  • 12.Soons JAM, Danckaers F, Keustermans W, et al. 3D morphometric analysis of the human incudomallear complex using clinical cone-beam CT. Hear Res. 2016;340:79–88. [DOI] [PubMed] [Google Scholar]
  • 13.Unur E, Ülger H, Ekinci N. Morphometrical and morphological variations of middle ear ossicles in the newborn. Erciyes Med J. 2002;24(2):57–63. [Google Scholar]
  • 14.Jin A, Xu P, Qu F. Variations in the labyrinthine segment of facial nerve canal revealed by high-resolution computed tomography. Auris Nasus Larynx. 2018;45(2):261–264. [DOI] [PubMed] [Google Scholar]
  • 15.Nager GT, Proctor B. Anatomie variations and anomalies involving the facial canal. Otolaryngol Clin North Am. 1991;24(3):531–553. [PubMed] [Google Scholar]
  • 16.Sayit AT, Gunbey HP, Sağlam D, Gunbey E, Kardaş Ş Çelenk Ç. Association between facial nerve second genu angle and facial canal dehiscence in patients with cholesteatoma: evaluation with temporal multidetector computed tomography and surgical findings. Braz J Otorhinolaryngol. 2019;85(3):365–370. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Redleaf MI, Blough RR. Distance from the labyrinthine portion of the facial nerve to the basal turn of the cochlea temporal bone histopathologic study. Ann Otol Rhinol Laryngol. 1996;105(4):323–326. [DOI] [PubMed] [Google Scholar]
  • 18.Abdelaziz AA. The role of preoperative computed tomography of temporal bone in atticotomy as a new tool for determining the approach. Indian J Otolaryngol Head Neck Surg. 2019;71(suppl 2):1272–1275. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Fauser J, Stenin I, Bauer M, et al. Toward an automatic preoperative pipeline for image-guided temporal bone surgery. Int J Comput Assist Radiol Surg. 2019;14(6):967–976. [DOI] [PubMed] [Google Scholar]
  • 20.Becker M, Kirschner M, Sakas G. Segmentation of risk structures for otologic surgery using the Probabilistic Active Shape Model (PASM). In: Yaniv Ziv R., Holmes David R. III, eds. Medical Imaging 2014: Image-Guided Procedures, Robotic Interventions, and Modeling, San Diego, CA, 15-20 February, 2014. SPIE; 2014. [Google Scholar]
  • 21.Neves CA, Tran ED, Kessler IM, Blevins NH. Fully automated preoperative segmentation of temporal bone structures from clinical CT scans. Sci Rep. 2021;11(1):116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Nikan S, Van Osch K, Bartling M, et al. PWD-3DNet: a deep learning-based fully-automated segmentation of multiple structures on temporal bone CT scans. IEEE Trans Image Processing. 2021;30:739–753. [DOI] [PubMed] [Google Scholar]
  • 23.Juliano AF. Cross sectional imaging of the ear and temporal bone. Head Neck Pathol. 2018;12(3):302–320. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES