Cell Instance Segmentation Dataset

This project contains the data for the paper: A. Bouyssoux, R. Fezzani and J. -C. Olivo-Marin, "Cell Instance Segmentation Using Z-Stacks In Digital Cytopathology," 2022 IEEE International Symposium on Biomedical Imaging (ISBI), 2022. The code associated with this project is available at: https://gitlab.com/vitadx/articles/zstacks_cell_instance_segmentation A new large Cell Instance Segmentation Dataset (CISD) is introduced. It comprises 3911 samples containing at least two touching or overlapping urothelial cells. Cell instances were manually annotated by trained cytotechnicians. All samples are extracted from 30 digital cytology slides stained with nine variations of Papanicolaou staining. The cytology slides are prepared from urine samples from healthy patients, using a Hologic ThinPrep®5000 processor, and routinely stained with the Agilent Dako CoverStainer®. The slides are finally digitized using a Hamamatsu NanoZoomer®S360 with 21 focal planes and centered on the best focus plane determined by the scanner autofocus. Note that cell instances with a bounding box width/height smaller than 10% of the sample width/height, as well as red blood cells and neutrophil cells were automatically filtered out because under-represented in the available data. Each sample is considered in three different manners in the CISD, allowing experimentation with different methods for handling Z-stack data and comparison with simple 2D acquisition: Center slice: the best focus plane only as determined by the scanner, which is equivalent to a 2D slide acquisition. Extended Depth of Field (EDF): the 21 planes merged in an image where all textured parts appear in focus. Raw Z-stack: the volume composed of the 21 focal planes. Once extracted, the dataset folder contains three subfolders, one for each of the three types of samples described here above, containing the images and stack of images. A JSON file contains the instance masks, encoded in RLE format.