Instance Segmentation of 2D Label-Free Microscopic Images using Deep Learning

Abstract

The precise detection and segmentation of cells in microscopic image sequences is an essential task in biomedical research, such as drug discovery and studying the development of tissues, organs, or entire organisms. However, the detection and segmentation of cells in phase contrast images with a halo and shade-off effects is still challenging. Lately, Mask Regional Convolutional Neural Network (Mask R-CNN) has been introduced for object detection and instance segmentation of natural images. This study investigates the efficacy of the Mask R-CNN to instantly detect and segment label-free microscopic images. The dataset used in this paper is taken from the ISBI cell tracking challenge. The Mask R-CNN is trained using different hyperparameters and compared to the U-Net model. Experimental results show that the Mask R-CNN model achieves 91.6 % when using ResNet-50 backbone and COCO weights.