Automatic cell counting via recognition of morphological characteristics

Abstract

Cell segmentation faces many challenges due to the diversity in cell shape and size. The different types of microscopy images do not help either. These two factors, combined together, make a one-size-fits-all algorithm very difficult to achieve. This project aims to develop a system whereby the cells in a given image may be counted automatically. It provides a review of some major works that look at convolutional neural network based methodologies as well as classical approaches of cell segmentation, for instance the Watershed Transform.

A dataset of fluorescent microscopy images has been obtained from the Centre for Molecular Medicine and Biobanking at the University of Malta, more specifically, images of Green Fluorescent Protein (GFP) transfected cells. GFP is a common tool used as a biological marker in molecular biology, medicine and cell biology. This dataset lacked training labels and, therefore, manual annotation of each individual cell was required.

This project makes use of an adaptation of the U-Net convolutional neural network to perform segmentation. Training images and labels were passed to the network and, after training the model, the network was able to produce a segmented image. Other image processing techniques, such as mathematical morphology, were also implemented to further improve the segmentation and, consequently, the cell count of the image.

The count of the test images was then compared with that of the ground truth images, taking account of true positives, false positives and false negatives. From these values, measurements such as precision, recall and f-measure were calculated and discussed. A number of experiments were conducted to determine which configuration delivered the best results. The best percentage error acquired was approximately 17% below the desired count.