Annotation procedure

To qualify for annotation, cells were required to have a clear and definable central feature, either a well-stained cuticular plate or a stereocilia bundle, connected to a labeled cell body of the expected shape and size for that tissue type/preparation. Most of the images within this collection were of hair cells with relatively normal appearance, with some instances of missing hair cells. Many of the annotated images are maximum projections of serial images acquired at multiple focal planes, often with the cytosol of one cell occluding another rendering the cytosol as an unreliable feature for bounding box detection. Thus, each hair cell annotation box bounds the cuticular plate enclosing the stereocilia bundle and is assigned a class label of cell type (IHC or OHC).

Bounding-box annotations were created and corrected in the labelImg open-source software²² and the HCAT open source software. The annotations were generated using a “human-in-the-loop”²³ paradigm as outlined in Fig. 1. Initial candidate annotations were generated using a neural network, which were then audited and refined to ensure every cell classification label was correct, and that bounding boxes tightly enclose the cell’s cuticular plate and stereocilia bundle without including any additional features. Since the images within the dataset were annotated by several observers, each image was then reviewed by a single lead annotator to ensure accuracy and minimize inter-annotator biases. Data annotations were saved as a separate xml file in the coco format with an identical filename to the associated image^24,25. Although uncommon, some images within the dataset were collected following application of certain insults and may include hair cells with damaged stereocilia bundles.