Eosinophilic Granulocyte Recognition on Whole Slide Imaging Using CircleSnake Deep Learning Model (DSI-SRP)
This DSI-SRP fellowship funded Yilin Liu to work in the laboratory of Dr. Yuankai Huo in the Department of Computer Science during the summer of 2022. Yilin is a senior with majors in Computer Science and Mathematics.
Eosinophil granulocytes, also called eosinophils or eosinophiles, are white blood
cells that are one of the immune system components responsible for combating
multicellular parasites and certain infections. Eosinophilic granulocyte is found in the
blood or tissues of patients with allergic diseases such as asthma, allergic rhinitis and
atopic dermatitis. The expression of clinical symptoms is generally associated with the
number of eosinophils in the diseased tissue. It was originally assumed that the
eosinophil served an exclusively protective role, such as defending against parasites in
host. More recently, the eosinophil is found to be active in inflammation that can
mediate allergic disease. The existing eosinophil identification is mainly relied on
human, which is easily confused with other types of cells such as Papilla Eosinophil and
Red Blood cell. The recognition of eosinophil also lacks a standardized evaluation
metric that is susceptible to human subjective error, especially from those inexperienced
In this study, Yilin utilized CircleSnake, an end-to-end circle contour deformation-
based segmentation method effective for segmented ball-shaped medical objects. The
trained model was able to recognize and distinguish between four different classes of
cells: the Eosinophil cell, the Papilla Eosinophil cell, the Red Blood cell, and the Red
Blood Cell Cluster cell.
Yilin manually annotated 50 WSIs into four class under the help of Dr. Wang. She also annotated 50 WSIs, including thousands of annotations, using QuPath software. After individual contour masks were generated from QuPath software, Yilin converted them into a standardized COCO format, a specific JSON structure dictating how labels and metadata are saved for an image dataset, where each contour is stored with different IDs and different class (Eos, Papilla Eos, RBC, and RBC Cluster) and different WSIs is also separated by different image IDs. She successfully set up the CircleSnake environment and reproduced the CircleSnake code using the original MoNuSeg dataset.
In addition to receiving support through a DSI-SRP fellowship, this project was supported and facilitated by the DSI Data Science Team through their regular summer workshops and demo sessions.