Usiigaci: Instance-aware cell tracking in stain-free phase contrast microscopy enabled by machine learning

Abstract

Stain-free, single-cell segmentation and tracking is tantamount to the holy grail of microscopic cell migration analysis. Phase contrast microscopy (PCM) images with cells at high density are notoriously difficult to segment accurately; thus, manual segmentation remains the de facto standard practice. In this work, we introduce Usiigaci, an all-in-one, semi-automated pipeline to segment, track, and visualize cell movement and morphological changes in PCM. Stain-free, instance-aware segmentation is accomplished using a mask regional convolutional neural network (Mask R-CNN). A Trackpy-based cell tracker with a graphical user interface is developed for cell tracking and data verification. The performance of Usiigaci is validated with electrotaxis of NIH/3T3 fibroblasts. Usiigaci provides highly accurate cell movement and morphological information for quantitative cell migration analysis.Cell migration is a fundamental cell behavior that underlies various phys-2 iological processes, including development, tissue maintenance, immunity, 3 and tissue regeneration, as well as pathological processes such as metastasis. 4Many in vitro as well as in vivo platforms have been developed to investigate 5 molecular mechanisms underlying cell migration in different microenviron-6 ments with the aid of microscopy. To analyze single-or collective-cell migra-7 tion, reliable segmentation of each individual cell in microscopic images is 8 necessary in order to extract location as well as morphological information. 9Among bright-field microscopy techniques, Zernike's phase contrast mi-10 croscopy (PCM) is favored by biologists for its ability to translate phase dif-11 ferences from cellular components into amplitude differences, so as to make 12 the cell membrane, the nucleus, and vacuoles more visible [1]. However, 13PCM images are notoriously difficult to segment correctly using conven-14 tional computer vision methods, due to the low contrast between cells and 15 their background [2]. For this reason, many cell migration experiments still 16 rely on fluorescent labeling of cells or manual tracking. Fluorescent labeling 17 of cells requires transgenic expression of fluorescent proteins or cells tagged 18 with fluorescent compounds, both of which can be toxic to cells and which 19 require extensive validation of phenotypic changes. Although thresholding 20 fluorescent images is relatively straightforward, cells that are in close prox-21 imity are often indistinguishable in threshold results. On the other hand, 22 manual tracking of cell migration is labor-intensive and prone to operator 23 error. Conducting high-throughput microscopy experiments is already possi-24 ble thanks to methodology and instrumental advances, but current analytical 25 techniques to interpret results quantitatively face major obstacles due to im-26 perfect cell segmentation and tracking [3]. Moreover, cell movement is not 27 the only parameter of interest in cell migration. For cell migration guided 28 by environmental gradients, shear stress, surface topolog...

Label-free imaging of immune cell dynamics in the living retina using adaptive optics

“…A customized semi-automated deep-learning based cell tracking software was employed to track and quantify the migration behaviour of cells. The software consisted of a deep learning-based cell detector with an encoder-decoder U-Net backbone architecture [12,21]. The U-Net was trained on phase-contrast AOSLO images with the centroids of 387 cells manually identified by expert graders obtained from four mice at either 6 or 24-hours post LPS injection.…”

“…The cell counts of a video were calculated by averaging the number of detected objects in the first 25 frames (5 seconds). To track the cells, centroid positions detected by the U-Net in adjacent frames were linked with a nearest neighbor search algorithm [21]. The deep learning strategy facilitated tracking of a large number of cells across multiple frames captured at different time points across inflammation.…”

Automated deep lineage tree analysis using a Bayesian single cell tracking approach

“…Accurate tracking of individual cells, and reconstruction of cell lineages from raw microscopy images remains a major bottleneck and rate-limiting step in subsequent data analysis despite major efforts in this area [17][18][19][20][21][22][23][24][25][26][27][28][29] . To extract multigenerational lineages, it is often necessary to manually annotate ancestor/descendant relationships in graphical lineage tree representations.…”

Automated cell tracking using StarDist and TrackMate

“…The study of cell motility typically involves recording cell behavior, using live-cell imaging, and tracking their movement over time 1,2 . To enable the analysis of such data, various software solutions have been developed [3][4][5][6][7][8][9] . However, despite the availability of these computational tools, manual tracking remains widely used among researchers due to the difficulty in setting up fully automated cell tracking analysis pipelines.…”

“…However, we acknowledge that using brightfield images may not always work directly with our pipeline, especially if cells display complex and interchanging shapes since StarDist is mostly designed to detect round or compact shapes. In this case, other tools, such as Usiigaci, could also be considered 8 . Still, brightfield images could also be artificially labeled using deep learning, transforming the brightfield dataset into a pseudo-fluorescence one, as can be done with ZeroCostDL4Mic already 15 .…”