Deep Learning-based Detection of Overlapping Cells

Abstract

Detection and quantification of cells in general is one of the key challenges in many clinical trials for disease diagnosis and monitoring. Automation of this task enables quantitative analysis of digital images with a high processing rate, which is a support to pathologist at various kind of analyses. Recent studies have already indicated that deep learning usually yield superior accuracy in the field of digital pathology. One of the challenges tackled by the researches is to detect cells in images when cells are highly overlapped, over illuminated or partially occluded with the noise. Therefore, we focused on two conceptually different deep learning models, specifically U-Net and Mask R-CNN, in order to evaluate their capability and performance on the detection of overlapping cells. The dataset used in the study contains different types of images, possible observed under different lighting conditions, and the amount of target cells may range from tens to thousands, therefore the algorithm is required to be flexible enough.