BEATing CILIAted cells

The dataset contains 119 videos of beating ciliated cells which can be used for research on CBF estimation. The dataset is also already partitioned into training, validation, and test sets containing 89 (185), 18 (36), and 12 (25) videos (cells) respectively; frames from this video are already extracted, and augmented and each one comes with its annotation file. 

The videos were acquired by an expert cytologist using a Nikon Eclipse 600 (with 1000x magnification) microscope equipped with a camera model MD6iS (Sony IMX236 Sensor).

Each video exhibits a variable number of ciliated cells, varying from 1 to more than 5 (in contexts where many cells are amassed), while the average video duration in the dataset is 93 seconds, with an upper bound of 352 seconds and the shortest being only 3 seconds long.

In addition to the number of cells and video duration, the videos differ from each other in terms of frame rate with an average of 23.63 Fps and a standard deviation of ± 2.54 Fps, and in terms of resolution.

IMPORTANT NOTE: The annotation of the cells was carried out by experienced personnel according to their interpretation.

The acquisition and preparation of this dataset have required a lot of work without any remuneration. We provide it also free of charge, but we ask those who intend to use our dataset the courtesy to cite the following papers (thanks in advance):

• Dimauro G., Barbaro N., Camporeale M.G., Fiore V., Gelardi M., Scalera M., DeepCilia: automated, Deep Learning based engine for precise Ciliary Beat Frequency estimation, Biomedical Signal Processing and Control. https://doi.org/10.1016/i.bspc.2023.105808
• G. Dimauro, F. Girardi, M. Gelardi, V. Bevilacqua, e D. Caivano, «Rhino-Cyt: A System for Supporting the Rhinologist in the Analysis of Nasal Cytology», in Intelligent Computing Theories and Application, 2018, vol. 10955 LNCS, pp. 619–630. doi: 10.1007/978-3-319-95933-7_71
• G. Dimauro, F. Girardi, D. Caivano, e L. N. Colizzi, «Personal Health E-Record - Toward an enabling Ambient Assisted Living Technology for communication and information sharing between patients and care providers», in Ambient Assisted Living, 2019, vol. 544. doi: 10.1007/978-3-030-05921-7_39

FURTHER DETAILS

The dataset contains a folder called “video” with the 119 videos in mp4 format and a csv file named “cilia.csv” that specifies:

• Number of ciliated cells in the video;
• Framerate of the video;
• Number of frames in the video;
• Duration of the video (in seconds);
• Resolution of the video.

In the folder called “frames” are stored the frames of the video, already extracted and preprocessed.

After partitioning the videos into training, validation, and test sets containing 89 (185), 18 (36), and 12 (25) videos (cells) respectively, the frames from this video were converted to grayscale, resized to 640x640 resolution and then underwent a data augmentation step that applied one operation between shearing (both vertical and horizontal) and random rotations in the range of ± 180 degrees.

The “frames” folder thus contains two subfolders, one called “images” and one called “annotations”; both folders contain 3 more subfolders called “train”, “validation” and “testing” each one containing respectively 36287 – 7216 - 5340 frames/txt files.

Each frame in the “images” folder has its corresponding annotation file in the “annotations” folder, the correspondence between a frame and its annotation can be detected since the two share the same name.

An annotation txt file contains a row for each bounding box (in the YOLO format) in the frame. In each row there are 5 values separated by an empty space; the values respectively represent:

• The class of the bounding box (always 0 since only strias are annotated)
• The x and y coordinates of the center of the bounding box (normalized)
• The horizontal and vertical dimensions of the bounding box (normalized)