Synchrotron‐Based Deep Learning Network of the Inner Ear: Development and Expert Validation Authors: Ashley Micuda, Kyle Rioux, Luke Helpard, Munir Demir Bajin, Zahra Kassam, Sachin Pandey, Jaehyun Hwang, Griselda T Romero‐Sanchez, Manas Sharma, Lorne S Parnes, Hanif M Ladak, Sumit K Agrawal Published: 2026 ABSTRACT ABSTRACTObjectiveTo develop and externally validate a deep learning segmentation network capable of automatically segmenting the inner ear in preoperative clinical computed tomography (CT) scans across various resolutions and protocols.MethodsA deep learning‐based segmentation network was developed using 100 cadaveric specimens that were scanned with synchrotron‐radiation phase contrast imaging (SR‐PCI) and various clinical CT scanners. Different acquisitions, protocols, and augmentations were used to create a total of 4,784 paired SR‐PCI and clinical three‐dimensional datasets used for deep learning training and model development. Performance and accuracy of the network were assessed on a separate unseen dataset and externally validated against manual segmentations from seven individual domain experts (otologists and radiologists), the mean expert performance, and a simultaneous truth and performance level (STAPLE) consensus segmentation.ResultsThe network pipeline significantly outperformed each individual expert segmentation, the average of the expert segmentations, and the STAPLE consensus segmentation. Compared to the SR‐PCI ground truth data, the network achieved a Dice similarity coefficient of 0.922, a maximum absolute Hausdorff distance of 0.329 mm, and an average Hausdorff distance of 0.006 mm on cone‐beam CT and helical CT with resolutions as low as 625 μm.ConclusionThis is the first automated segmentation algorithm for the inner ear that has been shown to outperform segmentations from domain experts, establishing a new clinical gold standard.Level of EvidenceN/A.Keywords: deep learning, inner ear, segmentation, STAPLE, synchrotron imaging Introduction The inner ear is a complex structure that exhibits considerable morphological variability across individuals, which can have significant clinical implications [1, 2, 3, 4]. Clinical computed tomography (CT) scans are routinely acquired before surgical interventions to identify inner ear malformations and assess individual anatomy. Imaging the entire inner ear has utility in diagnosing and managing many otologic conditions. Accurate assessment of the semicircular canals (SCC) and vestibule can enhance the diagnosis of vestibular disorders and aid in identifying congenital anomalies [5], which influence treatment strategies for both auditory and balance‐related dysfunctions. However, its main utility is extracting the cochlear anatomy, which is vital for tailoring interventions for hearing loss such as cochlear implants (CIs). Detailed knowledge of cochlear anatomy, both pre‐ and postoperatively, is relevant for selecting appropriate CI electrode lengths [6], preserving residual hearing [7], and tuning CI frequencies to individual patient cochlear geometry [8, 9, 10]. Despite clinical importance, segmentation (i.e., delineation) of the inner ear from conventional CT scans remains challenging. Manually obtaining measurements is time‐consuming, susceptible to inter‐observer variability [11], and limited by the resolution of clinical CT scans [12]. Single linear measurements used for preoperative planning and postoperative tuning, such as A‐value and B‐value, have limited accuracy in predicting total cochlear duct length [10, 13]. The ability to accurately segment the entire inner ear from CT scans could enhance precision of both linear and nonlinear measurements, including cochlear duct length estimation, used for surgical planning and tonotopic mapping. Many automated segmentation techniques have been proposed for temporal bone structures and in the context of CIs. Active‐shape model segmentation algorithms have become the most widely‐applied approach [14, 15, 16] for CIs in addition to machine learning models [17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30]. However, these methods have been developed using limited micro‐CT (μCT) datasets resulting in reduced generalizability for segmenting fine cochlear anatomy, such as the cochlear apex, and may not sufficiently capture the range of morphological variability in the human population. Machine learning has been successfully applied to many medical imaging tasks including segmentation, localization, and classification of anatomical structures. Deep learning, a subset of machine learning, commonly refers to the training of multi‐layered (deep) artificial neural networks. Convolutional neural networks (CNNs) in particular have advanced the application of deep learning for image processing tasks [22, 31, 32, 33, 34, 35] and have been adapted for semantic segmentation, where networks produce pixel‐ or voxel‐wise label maps matching the input image dimensions [32, 33]. CNNs have been applied to segment cochlear anatomy in clinical CT scans, but progress has been hindered by the scarcity of annotated clinical datasets and the use of two‐dimensional (2D) networks on inherently three‐dimensional (3D) structures [21, 22, 23, 24, 28]. The objective of the present work was to train and test a deep learning segmentation network capable of automatically segmenting the inner ear in preoperative clinical CT scans with various resolutions. To develop this segmentation network, a large dataset of paired high‐resolution synchrotron‐radiation phase contrast images (SR‐PCI) and a variety of clinical‐resolution CT scans of cadaveric inner ears was utilized. Finally, the accuracy of the network was externally validated by comparing against manual expert segmentations from otologists and radiologists. Materials and Methods Image Acquisition and Segmentation One hundred human cadaveric specimens (53 left, 47 right; 52 female, 41 male, 7 with unavailable sex data) were used for this study. The specimens captured a wide range of anatomical morphology, as quantified by the total inner ear volume (mean: 196 mm3; range: 143–257 mm3). All specimens were obtained as part of the Body Bequeathal Program at Western University, London, Ontario, Canada in accordance with the Anatomy Act of Ontario and Western's Committee for Cadaveric Use in Research (approval #122611). All specimens were scanned with clinical‐resolution CT and SR‐PCI, an imaging technique which achieves high‐contrast 3D images of cadaveric specimens using high‐energy X‐ray without the use of staining [36]. The deep learning model was developed using 84 specimens for the training dataset and 10 specimens for the internal validation dataset within the deep learning pipeline. Internal validation refers to the process of evaluating the model's performance on a separate dataset during deep learning training for tuning the model's hyperparameters. The remaining six specimens were used as an unseen test dataset for testing the accuracy of the model and for external validation against expert segmentations. The test set number was chosen to balance the time burden imposed on clinicians and manual segmentation quality. External validation will refer to expert segmentation of the test dataset hereafter. For the deep learning network to train on clinical data acquired in practice, the same specimens were scanned with a variety of clinical scanners to encompass different acquisitions, protocols, noise, and resolutions. The data used to develop the segmentation network is summarized in Table 1. TABLE 1Description of the 100 cochleae scanned with SR‐PCI and various clinical‐resolution scans. SR‐PCI data were used as ground truth and clinical‐resolution scans were used for deep learning training and testing. A single value for the resolution indicates isotropy and three values indicate anisotropy.Usage# of samples scannedModality/scanner typeResolution/protocolGround truth100SR‐PCI0.008–0.009 mmTraining/internal validation88Xoran xCAT CBCT0.1 mm880.3 mm40GE helical CT 0.488 × 0.488 × 0.625 mm, Soft tissue protocol 40 0.488 × 0.488 × 0.625 mm, Temporal bone protocol 89Canon helical CT 0.591 × 0.591 × 0.5–0.789 × 0.789 × 0.5 mm, Soft tissue protocol 89 0.591 × 0.591 × 0.5–0.789 × 0.789 × 0.5 mm, Temporal bone protocol 82Siemens angiography CT 0.48 mm, Natural protocol 82 0.48 mm, Sharp protocol Testing/external validation3Xoran xCAT CBCT0.3 mm3GE helical CT 0.234 × 0.234 × 0.625 mm, Temporal bone protocol Description of the 100 cochleae scanned with SR‐PCI and various clinical‐resolution scans. SR‐PCI data were used as ground truth and clinical‐resolution scans were used for deep learning training and testing. A single value for the resolution indicates isotropy and three values indicate anisotropy. 0.488 × 0.488 × 0.625 mm, Soft tissue protocol 0.488 × 0.488 × 0.625 mm, Temporal bone protocol 0.591 × 0.591 × 0.5–0.789 × 0.789 × 0.5 mm, Soft tissue protocol 0.591 × 0.591 × 0.5–0.789 × 0.789 × 0.5 mm, Temporal bone protocol 0.48 mm, Natural protocol 0.48 mm, Sharp protocol 0.234 × 0.234 × 0.625 mm, Temporal bone protocol The inner ear for each specimen was manually segmented on the SR‐PCI data and a consensus interpretation was achieved by three observers. The segmentations were used as ground truth labels for the corresponding clinical‐resolution CT scans of the same specimens (Figure 1). All scans were registered to each other in a clinical view using rigid landmark registration of clinical scans to the SR‐PCI reference; then intensity‐based rigid registration with the Brains module in 3D Slicer [37]. All input scans used for network training, validation, and testing were clinical‐resolution acquisitions. All scans were resampled to the same resolution and cropped using a bounding box around the region of interest. Cross‐sectional slice through the inner ear annotated with a segmentation derived from SR‐PCI data. (A) SR‐PCI slice with the corresponding segmentation, (B) clinical 300 μm cone‐beam CT slice with the aligned SR‐PCI segmentation overlayed, and (C) a 3D rendering of the SR‐PCI segmentation. Augmentations based on accepted basic image manipulations [32, 38] were performed to increase the dataset size and network robustness to variations in scans by creating realistic modifications of each clinical‐resolution scan. A copy of each clinical scan was generated by flipping the voxel coordinates along the mediolateral axis. This augmentation doubled the image set and produced a version of each sample representing the contralateral ear. Two additional versions of each scan were also generated with two levels of additive Gaussian noise. Lastly, an additional version was generated with blurring using a Gaussian filter across all dimensions [22]. An example of the image noise and blur augmentations is illustrated in Figure 2. Comparison of clinical‐resolution CT slices through the inner ear with varying levels of additive Gaussian noise for image augmentation. (A) Original clinical‐resolution CBCT scan, (B) clinical‐resolution CBCT scan with Gaussian noise with a small standard deviation, (C) clinical‐resolution CBCT scan with Gaussian noise with a large standard deviation, and (D) clinical‐resolution CBCT scan with Gaussian blurring across all dimensions. Overall, with the different acquisition, protocols, and augmentation, a total of 4790 paired synchrotron‐clinical 3D datasets were used in this study. For model development, including deep learning training and internal validation, 4784 paired scans were used. Training, validation, and test data were split using a random seed, with the requirement that bilateral cochleae be in the same dataset and all cochleae in the test set be unique. For the unseen test set, used for model testing and external validation, six paired scans were used. This provided ample data for deep learning and ensured that the cochlear variability present in the human population was well represented. Segmentation Network Training The NVIDIA MONAI platform (v1.4.0) [39] was used for developing and testing the deep learning network. A SegResNet architecture adapted from the work of Myronenko [40] was used as the network architecture in this study (Figure 3). The input volume size to the network was 304 × 304 × 304 voxels. The network was trained with a batch size of one, the Adam optimizer [41] with a learning rate of 0.0001, and a hybrid loss function which was a weighted sum of focal loss [42] and Dice loss [43, 44]. Convolutional neural network architecture adapted from Myronenko [40] utilized for automatic 3D segmentation of the inner ear. Segmentation Network Testing Comparison to SR‐PCI Ground Truth Segmentations To evaluate the performance and accuracy of the trained segmentation network, objective evaluation metrics were calculated on the unseen test set containing six cochleae with clinical scans of various resolutions, as shown in Table 1. The Dice score coefficient (DSC) [44] and Hausdorff distance (HD) [45] were computed for the network segmentation of the inner ear relative to the reference SR‐PCI segmentation for each test set scan. The DSC evaluates the spatial overlap between volumes, where a value of 1 indicates perfect overlap and 0 indicates no overlap. The maximum Hausdorff distance (HDmax) measures the maximum distance between the boundary points of the predicted and ground truth volumes in mm, capturing the worst‐case boundary error. The average Hausdorff distance (HDavg) measures the mean of the minimum distances in mm, reflecting the overall segmentation quality [46]. Lower HD values indicate higher boundary similarity. Comparison to Expert Manual Segmentations To perform real‐world external validation of the network, the resultant automatic segmentations from the network were compared to expert segmentations of the same clinical‐resolution cadaveric scans. Seven domain experts, including four radiologists and three otologists (with clinical experience ranging from 5–40 years), manually segmented the clinical‐resolution cadaveric CT test set using 3D Slicer (version 5.4.0, https://www.slicer.org/), an open‐source medical imaging software [47]. Experts were instructed to segment the entire inner ear and were able to use all segmentation tools available in the software including thresholding and 3D sphere brush. To avoid bias in the learning curve for segmenting with the software, all experts completed the segmentations in different orders. The expert segmentations were then up‐sampled to match the output spacing of the deep learning network. Taubin smoothing and island removal were applied as post‐processing steps to the final expert segmentations [48]. The simultaneous truth and performance level estimation (STAPLE) algorithm estimates the accuracy of experts and weighs the contribution of their segmentation to the final consensus segmentation accordingly [11]. In this work, the STAPLE algorithm was used to combine the seven expert segmentations for each scan to create a weighted “average” segmentation for comparison. Additionally, the individual expert scores were also averaged across the full test set for comparison to the deep learning and STAPLE algorithm across all evaluation metrics. Statistical Analysis Statistical analyses were performed for all objective metrics between the average expert score, the STAPLE segmentation, and the deep learning segmentation. A two‐sided Wilcoxon signed rank test was performed to compare the DSC, HDmax, and HDavg of the average expert segmentation scores, the STAPLE consensus segmentation, and the automatic deep learning segmentation network output. Additionally, a Kruskal–Wallis test was performed to compare all objective metrics for all the expert segmentations. The significance level was set at α < 0.05. All reported p‐values were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) correction. Results One hundred specimens with 4790 paired SR‐PCI and clinical scans were used for the development, model testing, and external validation of the deep learning segmentation pipeline. Qualitatively, the deep learning segmentations were very similar to the ground truth data. Figure 4 shows a segmentation contour for one unseen low‐resolution clinical test scan from the deep learning network (blue) and from the SR‐PCI ground truth (red). The derived contours are overlaid onto the paired synchrotron scan (Figure 4A) and the input clinical CT slice (Figure 4B). Surface renderings compare the deep learning network against the SR‐PCI ground truth (Figure 4C). Visualization of automatic deep learning segmentation from a low‐resolution clinical CT scan in comparison to the SR‐PCI ground truth segmentation. The segmentation output from the deep learning network is shown in blue and the SR‐PCI segmentation is shown in red. Both segmentation contours are overlayed on a cross‐sectional slice of (A) the SR‐PCI scan and (B) the input clinical scan. (C) A 3D surface rendering of both segmentations are shown overlaid. The 3D‐rendered manual expert segmentations, the STAPLE consensus algorithm, the proposed automatic deep learning segmentations, and the SR‐PCI ground truth segmentations are presented for the entire test set of six specimens in Figure 5. The experts are ordered according to average DSC, from lowest to highest. Qualitatively, the deep learning segmentation appeared the smoothest and most similar to the ground truth segmentation. Additionally, the deep learning segmentation was able to clearly capture the cochlear tip, which is used to calculate the total number of cochlear turns. Expert segmentations did not adequately capture the number of turns and missed parts of the SCC. The STAPLE segmentation was also unable to capture the number of cochlear turns but did achieve full segmentation of the SCC in all cases. Volume renderings of the manual expert segmentations (green), the STAPLE consensus segmentation (grey), the proposed deep learning segmentation (blue), and the ground truth SR‐PCI segmentations (red) for the unseen test set. The top three rows correspond to the helical CT test set scans with a resolution of 0.234 × 0.234 × 0.625 mm, and the bottom three rows correspond to the CBCT test set scans with a resolution of 0.3 mm isotropic. For all three metrics, the automatic segmentation pipeline performed the best, followed by the STAPLE consensus segmentation method, and the expert average performed the worst, as illustrated in Figure 6. The mean DSC was significantly higher for the proposed network pipeline (DSC = 0.922), compared with the STAPLE segmentation method (DSC = 0.882, p < 0.05) and the average expert score (DSC = 0.838, p < 0.05). The deep learning pipeline also had a significantly smaller average and maximum Hausdorff distance (HDavg = 0.006 mm, HDmax = 0.329 mm), compared to STAPLE (HDavg = 0.023 mm, p < 0.05; HDmax = 1.323 mm, p < 0.05), and the average expert score (HDavg = 0.028 mm, p < 0.05; HDmax = 1.334 mm, p < 0.05). A comparison between all three segmentation methods divided by scan type is presented in Table 2. For all segmentation methods, the CBCT scans achieved better scores (higher for DSC and lower for HD) compared to helical CT scans. For both helical CT and CBCT, the deep learning network achieved a HDmax of approximately 1 voxel error for the worst‐case error. Conversely, the STAPLE method and the average expert HDmax error was approximately 2–4 voxels. Boxplot comparison of individual expert segmentation scores for (A) DSC, (B) HDmax, and (C) HDavg in the first column. Boxplot comparison of the expert average score, the consensus STAPLE segmentation, and the deep learning network relative to the SR‐PCI ground truth segmentations for (D) DSC, (E) HDmax, and (F) HDavg. DSC, Dice score coefficient; HDavg, average Hausdorff distance; HDmax, maximum Hausdorff distance. Asterisks indicate statistical significance level where *p < 0.05. TABLE 2Segmentation results comparing the automatic deep learning segmentation, STAPLE algorithm, and expert average to the corresponding ground truth SR‐PCI segmentations for the three helical CT and three CBCT test set scans. Values are shown as mean [minimum, maximum].MethodScansDSCHDavg (mm)HDmax (mm)Automatic segmentationHelical CT test scans0.915 [0.881, 0.937]0.007 [0.004, 0.013]0.395 [0.320, 0.512]CBCT test scans0.930 [0.911, 0.956]0.005 [0.002, 0.007]0.263 [0.200, 0.320]STAPLEHelical CT test scans0.821 [0.677, 0.890]0.026 [0.019, 0.036]1.409 [1.160, 1.709]CBCT test scans0.865 [0.839, 0.981]0.019 [0.015, 0.024]1.237 [1.180, 1.305]Expert averageHelical CT test scans0.814 [0.677, 0.894]0.033 [0.013, 0.096]1.476 [0.650, 3.84]CBCT test scans0.849 [0.790, 0.887]0.022 [0.012, 0.042]1.207 [0.851, 2.171]Abbreviations: DSC, Dice score coefficient; HDavg, average Hausdorff distance; HDmax, maximum Hausdorff distance. Segmentation results comparing the automatic deep learning segmentation, STAPLE algorithm, and expert average to the corresponding ground truth SR‐PCI segmentations for the three helical CT and three CBCT test set scans. Values are shown as mean [minimum, maximum]. Abbreviations: DSC, Dice score coefficient; HDavg, average Hausdorff distance; HDmax, maximum Hausdorff distance. For all three metrics, the expert average and STAPLE segmentation method were not significantly different. These results suggest that the automatic segmentation network can outperform both the average expert segmentations and an expert consensus segmentation generated using the STAPLE method for all three objective metrics evaluated in this study. The Kruskal–Wallis test found a statistically significant difference in DSC (χ 2(6) = 22.65, p < 0.01), HDmax (χ 2(6) = 16.35, p < 0.05), and HDavg (χ 2(6) = 23.83, p < 0.001) between the expert segmentations. Following the FDR correction, all three metrics remained statistically significant (DSC, p < 0.01; HDmax p < 0.05; HDavg p < 0.01). This suggests that expert segmentations are subject to considerable variation for the same scans based on the segmentation techniques each expert used. Conversely, the variation across samples in the deep learning network is quite low, suggesting better consistency in the segmentation accuracy. Discussion A deep learning approach was introduced for automatic segmentation of the inner ear from various clinical CT scans, utilizing the SegResNet CNN. To train and evaluate the network, a novel dataset containing a collection of paired SR‐PCI and clinical‐resolution CT scans was used. The network demonstrated high accuracy in automatic segmentation of the inner ear, outperforming manual expert segmentations, the average expert segmentation scores, and the consensus expert segmentation in comparison to the SR‐PCI ground truth segmentations of the same clinical scans. Accurate inner ear segmentation in clinical CT data has become increasingly relevant due to advancements in deep learning methods, their utility in clinical practice, and precision medicine. Overall volume can be discerned from inner ear segmentations which may be useful for determining the dosage for transtympanic delivery of drugs which is now becoming a routine treatment for those with sudden sensorineural hearing loss [49]. Reliable segmentation also enables detailed morphological analysis, specifically for the cochlea, improving the accuracy of in vivo cochlear duct length, angular length, and volume metrics [50, 51, 52]. These measurements can inform preoperative CI electrode array selection to predict insertion depths needed for optimal insertion angles [53], aid in patient‐specific frequency tuning [50, 51, 54, 55, 56], and assess implant location after cochlear implantation [57]. Additionally, segmentations can augment the information available to clinicians, particularly in the field of CIs, as robotic‐assisted surgery and trajectory planning is becoming globally available [58, 59]. Therefore, integrating segmentation into routine clinical workflows is vital for diagnosing and treating auditory and vestibular conditions. Deep learning networks have previously been applied to inner ear [19, 20, 22, 24, 25, 26, 27] and cochlear segmentation [21, 23, 30]. A recently published summary of the machine‐learning‐based methods for inner segmentation showed a range of DSC from 0.85 to 0.94 [19]. Study limitations included utilization of a 2D convolutional neural network architecture to perform segmentation [23]; omitting spatial context included in 3D clinical data; the use of ground truth data manually segmented on clinical scans [24], which are inherently limited to the quality of the scans and can suffer from interobserver variability (as shown in the current work); the use of μCT as ground truth [22, 25], which has lower resolution and contrast than SR‐PCI; and testing only on high‐resolution scans (< 0.15 mm isotropic) [19], which is not the typical clinical‐resolution available at most clinics. The current work achieved similar DSC and HD scores on much lower resolution scans (down to 625 μm). The current study aimed to overcome the challenges of the previous works by using SR‐PCI as ground truth data, which enabled gold‐standard ground truth segmentations to be paired with clinical‐resolution scans, allowing the network to learn consistent boundaries across various low‐resolution protocols. This resulted in highly accurate segmentations that were not influenced by blooming [60] into the membranous labyrinth due to the dense otic capsule and did not need additional smoothing or island removal. Additionally, the use of a large dataset captured the variability seen in the general population. Few studies have been able to use more than one resolution of clinical scans. In this study, multiple resolutions, manufacturers, and protocols of clinical scans were used to increase generalizability of the network. Lastly, prior inner ear studies using expert segmentations as the reference could not report model‐over‐expert performance benefits. In contrast, this study had gold‐standard SR‐PCI scans, which eliminated expert segmentation variability and provided a true reference segmentation for comparison. This uniquely positioned this study to compare deep learning segmentations versus expert segmentations against a true standard reference. Taken together, these advantages allow for this work to surpass the current gold standard of expert segmentations. Although all experts had greater than 5 years of experience, choosing different thresholds and segmentation techniques resulted in significantly more variation than the deep learning network across the test set. In addition, the deep learning network was able to discern the cochlear apex in all tested scans, which is useful when determining cochlear duct length [10] for CI array selection and tonotopic mapping. Objective metrics resulted in a HDmax of approximately 1 voxel and a HDavg of less than 1 voxel, even when tested on standard low‐resolution helical CT (625 μm) and CBCT (300 μm). This compares very favorably against previous studies. The proposed pipeline can be easily integrated into clinical practice. Current manual segmentations take approximately 10–20 min per scan by domain experts, who often lack the time to perform segmentations even with the use of semi‐automated tools. In contrast, the deep learning network takes approximately 1 min per scan to generate an accurate segmentation of the inner ear. Commercially available otologic planning software has demonstrated successful preoperative and postoperative modeling and planning [61], and future work may include augmenting these results with the deep learning segmentation. Furthermore, the development of deep learning models for intracochlear anatomy would be a valuable future direction and have utility across a wide range of applications. The integration of deep learning segmentation networks into clinical workflows in the field of otolaryngology has the potential to improve diagnostic and treatment capabilities such as anatomy‐based fitting after CI surgery, since many important anatomical structures and fine details are often not visible in current clinical‐resolution scans. The deep learning network presented here showcases the ability of deep learning techniques to extract patient‐specific information from routine imaging, supporting the field's shift toward precision medicine. Conclusion The deep learning network pipeline presented was shown to outperform all individual experts, the average of the experts, and the STAPLE segmentation on objective measurements. This network was developed using a large dataset of paired high‐resolution, high‐contrast SR‐PCI data and a variety of clinical‐resolution CT scans with multiple resolutions, manufacturers, and protocols. The use of SR‐PCI for network development enabled gold‐standard ground truth segmentations, allowing the network to learn consistent boundaries across various low‐resolution protocols. These innovations establish a new benchmark for inner ear segmentation, exceeding the current gold standard. Funding Part of the research described in this paper was performed at the Canadian Light Source, a national research facility of the University of Saskatchewan, which is supported by the Canada Foundation for Innovation (CFI), the Natural Sciences and Engineering Research Council (NSERC) (RGPIN‐2022‐04184), the National Research Council (NRC), the Canadian Institutes of Health Research (CIHR), the Government of Saskatchewan, and the University of Saskatchewan. A.M. acknowledges funding from the Queen Elizabeth II Graduate Scholarship in Science and Technology. H.M.L. acknowledges funding from the NSERC Discovery Grant program. Conflicts of Interest S.K.A. is on the Surgical Advisory Board of MED‐EL GmBH. MED‐EL GmBH has previously provided a donation to Western University and has supported MITACS grants. However, the study was completed independently; industry had no role in the study design; data collection, analysis, or interpretation; manuscript preparation; or the decision to submit for publication. This work was initiated while L.H. was a PhD student at Western University and was completed independently of his current employment at MED‐EL GmBH. Acknowledgments Lauren Siegel provided manuscript editing and review. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.