Segmentation of Nuclei in Microscopy Images Across Varied Experimental Systems

Abstract

Nuclei detection in microscopy images is a major bottleneck in the discovery of new and effective drugs. Researchers need to test thousands of chemical compounds to find something of therapeutic efficacy. Nucleus being the most prominent part of a cell helps in the identification of individual cells in a sample and by analyzing the cell’s reaction to various treatments the researchers can infer the underlying biological process at work. Automating the process of nuclei detection can help unlock cures faster and speedup drug discovery. In this paper, we propose a custom encoder–decoder style fully convolutional neural network architecture with residual blocks and skip connections which achieves state-of-the-art accuracy. We also use spatial transformations for data augmentation to make our model generalize better. Our proposed model is capable of segmenting nuclei effectively across a wide variety of cell types and experimental systems. Automated nuclei detection is projected to improve throughput for research in the biomedical field by saving researchers several hundred thousand hours of effort every year.