Brain Cell Segmentation and Detection from the LIVECell Dataset Using Deep Learning with the EfficientDet Model

Abstract

Neurological disorders are a common cause of death and disability in the world. One of the difficulties is that the response to treatment is hard to evaluate and quantify. Neuronal cells could be reviewed using light microscopy modality. However, it is challenging and time-consumption for detecting and segmenting even with experts. Therefore, accurate instance segmentation of cells is crucial. With the advancement of computer vision and deep learning, the detection and segmentation of cells with the aid of computers might lead to the resolution of these issues. This article presents methodological segmentation and brain cell markers derived from the LIVECell dataset. LIVECell dataset is a huge, high-quality, hand-annotated, skillfully constructed dataset including 1,686,352 individual cells in a variety of cell formats and density cultures. The proposed training model using the EfficientDet model is a model that can perform segmentation training and mark individual cell units by setting bounding boxes, which makes optimal use of LIVECell data without the need to use the single modeling two for two tasks. As a result, the intersection over union (IoU) is between 50.0% and 80.0%, with a mean accuracy precision(AP) of 58.824% and an average false-negative ratio (AFNR) of 56.566%, relying on the bounds segment bound box to mark separate objects with IoU values. This result may open up future functional studies on the individual kernel differences in performance and architecture of neural networks for segmentation implementation.