SVPath: A Deep Learning Tool for Analysis of Stria Vascularis from Histology Slides

The stria vascularis (SV) is a vascular structure adjacent to the lateral wall of the cochlea that plays a critical role in audition through the secretion of endolymph, maintenance of endocochlear potential, and formation of the blood-labyrinth barrier within the inner ear [1]. The structure primarily comprises three major cell types: marginal, intermediate, and basal. It also includes a capillary bed and other less common cells, such as spindle and endothelial cells [2‐5].

Since the 1900s, various otologic pathologies have been linked to dysfunction of the SV [2]. Carraro et al. and other groups have found marked atrophy of the SV and its capillary bed in patients with presbycusis, and additional studies have shown how morphologic changes of the SV occur with noise-induced and autoimmune hearing loss [3‐7]. Notably, in the mid-1900s, autonomic dysregulation linked to vasospasms in the stria and surrounding vasculature was linked to Meniere’s disease, and the structure is still thought to play a role in cytomegalovirus-induced neonatal hearing loss and sudden sensorineural hearing loss [3‐5, 8].

Researchers have assessed histologic slides of the inner ear from deceased humans and other mammals to understand how pathologic processes affect different microstructures for a century [9]. Most histological studies of the inner ear, including those of the SV and its capillary bed, have been done manually, annotating and segmenting these structures from slides [5, 7, 9, 10]. Manual annotation is time consuming and subject to inter-rater variability [10]. These practical constraints may hinder the discovery of SV-related changes associated with human disease.

In recent years, advancements in digital microscopy have facilitated the curation of whole-slide pathology imaging (WSI) databases. Researchers have begun applying deep learning techniques to these databases to automate the analysis of structures from histology slides [11‐13]. Analysis and segmentation of the SV from temporal bone WSI pose many practical challenges. WSI files are large, often in the order of 1–2 gigabytes per image. Deep learning methods applied to these large files are computationally expensive and slow to train and implement [11]. Their performance diminishes with inconsistent staining of the slides and insufficient training labels [12]. Further, deep learning methods can have difficulty accurately segmenting small structures like the SV within large slides. For example, in terms of pixels, the SV is less than a millionth of the entire WSI.

Our aim in this paper is to propose a deep learning tool, SVPath, for automatic analysis of the SV from WSI. To resolve some of the challenges outlined above, we split this task into three subtasks: detecting the SV within the WSI, segmenting the SV and its capillary bed, and extracting relevant metrics from the segmentation. We then validate these subtasks by applying this technique to slides from two ear WSI from different macaques.

Qualitative assessment of [YOLOv8]² and nnUnet training and validation is shown in Figs. 4 and 5, respectively. Applying both steps took an average of ~ 40 s per WSI. As mentioned above, the performance of [YOLOv8]², the object detection portion of the model, is measured by mAP, precision, and recall. These metrics were captured for detecting the cochlea and the SV for both macaque ear M1, the internal dataset, and M2, the external dataset, and presented in Table 1. On average, the mAP for cochlea and SV detection for the internal dataset was 97.5%, which fell slightly to 96.4% for the external dataset. The average precision and recall for both structures in the internal dataset were 94.1% and 96.8%, respectively, slightly higher than the external dataset (average precision/recall: 92.7%/95.6%). Overall, performance across all metrics was greater than 90% for both structures across the internal and external datasets.

Metrics to assess the performance of the nnUnet included average dice score and IoU. A score of 1 indicates a perfect overlap between the ground truth mask and the mask generated by nnUnet. Average SV segmentation was more accurate than segmentation of the CB in both internal (SV: 0.95, CB: 0.88) and external (SV: 0.84, CB: 0.75) validation datasets (Table 2). All the SV and most CB segmentations met the acceptable criteria dice score > 0.7 for both the internal and external datasets.

Figure 6 demonstrates how SVAT could be applied to compare two ears. The M1 and M2 macaque ears represent normal monkey ears stained and sectioned using a similar protocol. There was no statistical difference between M1 and M2 SV area, SV width, or area per capillary (p = 0.48, p = 0.34, and p = 0.42, respectively). However, a statistical difference was noted for nuclei per SV between the M1 and M2 (p < 0.001).

The stria vascularis is integral to normal cochlear physiology and may play a role in several otologic pathologies. However, further understanding of this structure has been in part limited by the shortcomings of manual segmentation and analysis of histopathological slides. The SVPath pipeline that we have developed leverages novel methods in deep learning to facilitate accurate and high-throughput extraction and analysis of the SV from temporal bone WSI.

The validation results for the [YOLOv8]² method demonstrate that using two separate YOLOv8 neural networks in series can successfully extract the location of small structures, such as the SV, from within a large WSI. A high degree of accuracy for detecting the cochlea from within the WSI and the SV from within the cochlea region was expected since using the YOLO neural network framework to detect objects has previously been validated [15]. Similarly, the validation results for nnUnet indicate that SVPath can successfully segment the SV and its capillary bed from patches extracted by [YOLOv8]² [16]. The difference in performance between SV and capillary bed relates to a few factors. Generally, neural networks for image segmentation perform better when more pixels represent a label relative to the background. Given that the SV is represented by significantly more pixels than the CB within any given patch, a decrease in performance is expected. Additionally, the capillary bed’s shape, size, and other features are more variable than the SV. Generating more training data to capture this variability may lead to better performance. While SV segmentation outperformed capillary bed segmentation, most segmented capillary beds still met the acceptable criteria (dice score > 0.7). Both the YOLOv82 and the nnUnet, when applied to an external dataset consisting of a separate macaque ear specimen not used for training and internal validation, achieved similar performance across both tasks, indicating that this method is generalizable.

Regarding the SV analysis, given that both the external and internal datasets were collected from normal ears, a lack of a statistically significant difference between the datasets for the metrics analyzed was expected. However, the number of nuclei was found to be different between the two ears. This may be due to the difference in size between the two datasets; only 20 SV were analyzed for the M2 ear compared to 220 SV for the M1 ear. The purpose of including SVAT results was to demonstrate how it can extract clinically relevant features from the SV. Future work analyzing more sections across various normal macaque ears is required to generalize these findings.

Successful validation of SVPath to analyze SV has clinical implications. As mentioned above, SV plays a role in many ear diseases; however, the pathophysiology remains poorly understood. For instance, contrary to convention, Wu et al. suggested that SV atrophy was a secondary effect of presbycusis rather than its cause [21]. Since then, other groups, including Lang et al., have suggested that SV may be involved early in the pathogenesis of presbycusis [22]. In both studies, the groups manually sampled the SV at a few sites per slide for analysis and used a single metric (SV cross-sectional area vs. SV thickness). By automating the analysis, SVPath facilitates large-scale, complete SV analysis that may lead to insights about the role of SV in presbycusis and other diseases. SVPath allows researchers to capture metrics to quantify SV atrophy in multiple ways and focus on the vasculature within the SV. Further, SVPath, by creating a standard method to analyze SV, avoids the pitfalls of inter-rater variability in manual segmentation.

Other groups have segmented and analyzed WSI using neural networks. Khened et al. proposed a generalized framework that used an ensemble of neural networks based on a U-net with various backbones to improve the model’s overall accuracy [12]. Guo et al. created quick yet accurate deep-learning methods to analyze breast images, where they used a patch-based method to extract cancerous regions and then evaluated these regions at different resolutions to gather course and fine features [13]. Both methods offer accurate WSI segmentation, but the SVPath pipeline offers two significant advantages. The proposed pipeline captures features at the level of the WSI, like the cochlea, and features at the level of the cochlea, like SV. This pipeline could easily extend to extract features ranging from relatively large structures like semicircular canals to finer structures like the tectorial membrane. Additionally, this pipeline preserves global location information of the SV relative to the original WSI, allowing researchers to overlay structures on the WSI. Using existing image registration methods, stacked, aligned WSI can generate 3D models of histologic structures from temporal bone slides. Such information could further lead to understanding of how complex spatial relationships contribute to various pathologies. There may be other proprietary deep learning pathology tools that function like SVPath; however, these tools can be financially prohibitive. SVPath is built on other open-source tools, thereby enabling increased access such technology.

There are some limitations of the SVPath tool. While SVPath was validated on an external dataset, this dataset was stained and sectioned similarly to the training data. Uniformity of staining for both datasets was ensured by manually inspecting each image processed. Adding and training with differently stained datasets or datasets sectioned in other planes in the future could further improve the generalizability of this tool. Additionally, this tool was limited to training and testing on healthy macaque ears; future work should focus on applying this technique to diseased ears and human temporal bone specimen WSI to improve the clinical utility. From a technical standpoint, training multiple neural networks was computationally expensive, requiring a dedicated GPU. Additionally, preserving images at multiple levels requires significant memory and storage. Despite these disadvantages, the proposed SVPath tool has the potential to improve current analysis techniques for the SV dramatically.

SVPath provides a deep learning pipeline for automatic analysis of SV. This pipeline can be extended to other structures within temporal bone WSI and facilitate 3D analysis of such structures. As the digital temporal bone WSI database grows, integrating such tools will enable researchers to conduct rapid, large-scale investigations of structures within temporal bone WSI. Such studies could translate to new insights into the mechanism of various otologic pathologies and guide future interventions.