Large-scale annotated dataset for cochlear hair cell detection and classification

Abstract

Our sense of hearing is mediated by cochlear hair cells, localized within the sensory epithelium called the organ of Corti. There are two types of hair cells in the cochlea, which are organized in one row of inner hair cells and three rows of outer hair cells. Each cochlea contains a few thousands of hair cells, and their survival is essential for our perception of sound because they are terminally differentiated and do not regenerate after insult. It is often desirable in hearing research to quantify the number of hair cells within cochlear samples, in both pathological conditions, and in response to treatment. However, the sheer number of cells along the cochlea makes manual quantification impractical. Machine learning can be used to overcome this challenge by automating the quantification process but requires a vast and diverse dataset for effective training. In this study, we present a large collection of annotated cochlear hair-cell datasets, labeled with commonly used hair-cell markers and imaged using various fluorescence microscopy techniques. The collection includes samples from mouse, human, pig and guinea pig cochlear tissue, from normal conditions and following in-vivo and in-vitro ototoxic drug application. The dataset includes over 90'000 hair cells, all of which have been manually identified and annotated as one of two cell types: inner hair cells and outer hair cells. This dataset is the result of a collaborative effort from multiple laboratories and has been carefully curated to represent a variety of imaging techniques. With suggested usage parameters and a well-described annotation procedure, this collection can facilitate the development of generalizable cochlear hair cell detection models or serve as a starting point for fine-tuning models for other analysis tasks. By providing this dataset, we aim to supply other groups within the hearing research community with the opportunity to develop their own tools with which to analyze cochlear imaging data more fully, accurately, and with greater ease.

Keywords: annotation; cochlea; detection; hair cells; inner hair cell; machine-learning-ready data; outer hair cell.

Figure 1:. Cochlear hair cells annotation workflow using a ‘human-in-the-loop’ annotation paradigm.

(A) An exemplar cochlea from our dataset with IHCs and OHCs annotated. Bounding boxes and classification labels were generated for all hair cells along the sensory epithelium. Representative regions in the (B) apex, (C) middle, and (D) base are shown as insets. Bounding boxes and labels were generated using labelImg (E), an open-source object annotation software, with green boxes annotating OHCs and yellow boxes annotating IHCs. The annotation procedure was optimized using the human-in-the-loop paradigm (F), where annotations were first generated by a preliminary neural network, then manually corrected and used to further train and improve the preliminary network, iteratively enabling more accurate candidate detections.