Automatic processing of microscopic images is a critical component in analysis of many biological experiments. In recent years, much effort was devoted to develop algorithms and automatic tools for this task. However, most algorithms are designed for fluorescence microscopy, thus bright-field microscopy, which demonstrates morphological alteration and is harder to process, has been neglected.
We suggest a new approach for multi-cellular analysis of bright field microscopy. The main idea is to use the natural textural information for both image segmentation and appearance-based classification tasks. MultiCellSeg is applied on DIC images from time-lapse wound healing experiments to verify that HGF/SF accelerates healing, and to demonstrate that the healing rate is linear both for treated and untreated cells. It is also applied as the first step in a texture-classification application to measure cell scattering, an approach that proved to be extremely accurate, achieving perfect agreement with manual expert's visual tagging.
MultiCellSeg applies classification to the task of image segmentation to cellular and background regions. To the best of our knowledge, this is the first attempt to apply Machine Learning to this problem and to conduct a comprehensive comparison of its performance with that of a segmentation algorithm designed for this purpose. Our approach surpasses the existing algorithms in performing this task for a wide range of scales, illumination conditions, and cell types without the need to tune parameters, which is critical in such applications.
MultiCellSeg's local-patches classification approach significantly surpasses TScratch's in all data sets but one (tscratch, see ). This data set was taken from the TScratch package, with many images that contain scattered cells. The region-classification is designed to deal with this problem. When considering the final segmentation, MultiCellSeg significantly tops the alternative in all data sets.
In principle, the second phase of MultiCellSeg may be plugged in to enhance the performance of Scratch's second phase, but TScratch seems to be less sensitive to small details, which results in significantly less fine regions then with our approach.
Utilization of several types of features on several scales makes MultiCellSeg robust for varying conditions. In contrast to other approaches that tend to refrain from fine details to avoid gross mistakes or use data-specific assumptions, our algorithm operates in higher spatial resolution, detects small regions of interest and then decides whether to keep or to discard them via post processing (regional classification), in a fully automated manner. As a result, in many images where the wound is almost healed, our algorithm performs satisfactorily, whereas other algorithms fail to mark open regions, as exemplified in .
To further enhance the proposed segmentation performance, one can suit a model to fit a specific experiment, cell type or imaging conditions. This can be exceedingly useful nowadays, when high-throughput experiments are performed, each with hundreds of images 
. To this end, one (or more) image(s) should be manually marked to apply the training phase in our algorithm. This process is only partly automatic, but it requires no-parameter setting and may result in notable improvement in performance with minimal effort.
The automatic, accurate zero-parameters MultiCellSeg may serve as a tool for various biological analyses. MultiCellSeg's Matlab source code is freely available as standalone software to allow others to use it for wound healing analyses, multi-cellular bright field cells segmentation, and for other applications yet to evolve. The source code and accompanying graphical user nterface (GUI) can be found at http://www.cs.tau.ac.il/~assafzar/MultiCellSeg.zip
, it is recommended to read carefully the README file (http://www.cs.tau.ac.il/~assafzar/MultiCellSeg_README
) before applying it. In the future, we plan to add training capabilities to enable specific designated models for different cell lines and imaging conditions and/or to integrate it as part of a larger project (e.g., 
Wound healing assay is common and is applied by many research groups, but its analysis is very narrow in the sense that only a few measures are considered: the healing rate is calculated over a short period of time. The approach presented here can become the cornerstone for novel methods to be exploited in wound healing analysis. To analyze large data sets such as frequently sampled wound healing assays (acquired by time lapse microscopy), we suggest to perform manual marking of a few images to train a classifier that will be used to segment the entire time-lapse experiment. Producing these high-temporal-resolution progress graphs may reveal biological processes that are currently unknown, such as the linearity of the healing process, as described here.
Another potential corollary is to model the motion patterns of cells throughout the healing process. This is an open question of current interest (e.g., 
). Modeling cellular motility patterns under stimulants/inhibitors treatments may facilitate the understanding of cell motility mechanisms and enable the development of new anti-metastatic drugs. A correct, high-throughput, partitioning to occupied and background regions can be the first step in developing such an analysis.
An additional application for bright field multi-cellular segmentation is in cell scatter assay. Image texture histogram of cellular regions is used to define the degree of scattering in an inherently different approach than prior attempts that focus on counting single and clustered cells. Information extracted from multi-cellular bright field microscopy can thus be used to distinguish between different molecular-related cellular motility and morphology phenomena.