| Computer Vision and Pattern Discovery for Bioimages |
| |
| The group of Computer Vision and Pattern Discovery for Bioimages focuses on applying advanced computer vision, machine learning and mathematical models to elucidate the complex behavior of biological systems. We analyze images from wide-field and confocal microscopes, including image data sets from high-throughput screens. The trend towards quantitative biology has spawned new areas of research, especially in the area of digital imaging where thousands of images are acquired automatically through robotic systems of chemical and cell assays handling. These images are then analyzed and used to create new biological hypotheses that are further validated using other experimental means. Our contributions to high throughput and high content imaging are to provide accurate and fast computational methods for the data mining of large image data sets. |
| |
| Research Projects |
| |
1. A Multi-resolution Stochastic Level Set Method for the Segmentation of Bioimages
Yan Nei Law (BII), Hwee Kuan Lee (BII), Andy M. Yip (NUS/Math) |
| |
Motivation:
Image segmentation is indispensable in bioimaging applications. Due to the complex nature of bioimages, optimization model based methods often give the best segmentation results, provided that the underlying optimization problem is solved accurately. The Mumford-Shah model is especially useful for many bioimages. But existing algorithms for this model require a good initial solution to obtain good results and are therefore impractical. To make the model practical, it is essential to develop an algorithm which can compute the global or near global optimal solution efficiently. |
| |
 |
| |
Results:
We propose a hybrid approach to optimize the Mumford-Shah energy for image segmentation. It solves the problem of sensitivity to the initial solution. We also propose a multi-resolution approach to reduce the computational cost and enhance the search for the global minimum. We derived a useful theoretical result relating solutions at different resolutions. The proposed method works very well for images with a high level of noise and wide spread intensity levels such as those from live cell microscopy. Moreover, it can detect objects without clear boundaries, objects with complex shapes and clusters of small objects. We validated the method using images obtained from microscopy and MRI. The results show that the algorithm can accurately extract useful structures such as cell nuclei, brain tumors and proliferation marker PCNA in the images. |
| |
 |
| |
| 2. Automatic and Quantitative Measurement of Neural Cell Outgrowths |
| |
Understanding the biology of neural cells is important in designing treatments for neural related diseases. A unique trait of neural cell is that it has neurites which connect to other neural cells. These neurites outgrowth is a fundamental characteristic of neurons and they eventually form synapses and proper functioning of the nervous system depends on the formation of proper connections.
In this study, a high throughput image screening system is developed. This system includes the microscopy setup as well as advanced computer vision applications to process the images. Our task in the computer vision and pattern discovery group is to develop fast and accurate software to process thousands of images generated through the high throughput microscopy system.
Our new method combines the level-set and watershed methods in a specific way to achieve fast and accurate segmentation of the neural cells.
As shown in the figures, neural cells have outgrowths that overlap and the cytoplasms of many cells touches each other. Many algorithms in the literature could not separate cells that overlap and touch each other. Our method is designed specifically to overcome this difficulty.
Our method performs much better than currently available software, the error rate of our method, validated against a set of about 6000 cells is 6.5% while the error rate for METAMORPH (http://www.moleculardevices.com/pages/software/metamorph.html) on the same set of data is 25.5%
|
| |
 |
| Top row shows two fluorescence microscopy images with crossing and overlapping neurites and cytoplasm. Bottom row shows successful segmentation of the neurites and cytoplasm using a combination of level-set and wateshed |
| |
| 3. Automated Protein Distribution Detection in Images from High-throughput Image-based siRNA Library Screens |
| |
The availability of RNA interference (RNAi) libraries, automated microscopy and computational methods enables millions of biochemical assays to be carried out simultaneously. This allows systematic, data driven high-throughput experiments to generate biological hypotheses that can then be verified with other techniques. Such high-throughput screening holds great potential for new discoveries and is especially useful in drug screening. In this study, we present a computational framework for an automatic detection of changes in images of in vitro cultured keratinocytes when phosphatase genes are silenced using RNAi technology. In these high-throughput assays, the change in pattern only happens in 1-2% of the cells and fewer than 1 in 10 genes that are silenced cause phenotypic changes in the keratin intermediate filament network. In this situation, a work-flow is required in identifying the one or two cells in an image that manifest phenotypic changes.
Measuring generic features on the whole image such as cell morphology, edge or haralick texture, will fail to correctly detect positive 'hits' in this assay.
We present a framework for applying computer vision on image data generated by high-throughput screening. Our approach differs from a general machine learning approach, in that we incorporate prior biological knowledge into our algorithm. Our detection method is based on phenotypic changes to keratin proteins fused to the green fluorescent protein (GFP). Upon broad-spectrum pharmacological inhibition of phosphatases, small keratin aggregates appear in cells in addition to the normal reticular network seen in untreated cells. We wish to identify the specific phosphatase responsible for this phenotypic change using an RNAi based approach to silence each phosphatase gene in turn.
Computational analyses performed on collections of images are used to optimize each step of our work flow,
resulting in a multi-step process that can successfully classify images of cells with normal reticular network from those images containing cells with the aggregation phenotype. We have taken a stepwise approach to the problem, combining different analyses, each of which solves a portion of the problem. These include aggregate enhancement, aggregate clustering, edge detection and circular object detection, prior to final classification. This strategy has been instrumental in our ability to successfully detect cells containing protein aggregates.
|
| |
 |
| |
| Images with mutants detected automatically and marked. Mutants exhibits spot like phenotypes in their Keratin proteins while the Keratin proteins of wild types forms a network structure. |
| |
| 4. Automated Nucleus and Cells Detection using a Region Based Ellipse Detector |
| |
Fluorescence microscopy uses the phenomenon of fluorescence to image biological objects such as certain targeted proteins in the cells. The protein of interest is tagged with a fluorescence material that excites a specific wavelength of light when exposed to a laser. The excited light is then collected in photo-multipler tubes or charge coupled devices. Fluorescence microscopy yields many beautiful images with objects of various shapes, depending on the protein of interest. To study the distribution of proteins, the location of the nucleus provides a reference point to the cell under study. Therefore, locating the nucleus accurately is vital in elucidating the protein distribution in a cell.
Nuclei, and even cells, are often elliptical in shape and therefore the problem of locating the nucleus or cell is often reduced to the problem of finding ellipses in bioimages.
To reduce the effects of photo bleaching and to reduce the stress on the cells under incident laser light, low laser power is usually preferred in fluorescent microscopy. Low incident light results in low contrast and noisy images.
Under such conditions, the state-of-the-art ellipse detector in the literature does not accurately detect the nucleus. In fact, most ellipse detectors rely on edge maps which depends on the accuracy of edge detectors. It is well known that all edge detectors fail under low contrast and high noise situation.
We developed a new ellipse detector that is region based and does not depend on edge maps. This ellipse detector is designed to deal with low contrast and high noise situation. Because it is region based, the noise gets averaged out. We have shown that our method works for fluorescence images where the popular
Randomized Hough Transform failed.
|
| |
 |
| Nucleus of breast cancer cells successfully detected even with very low contrast and high noise. |
| |
 |
| Red blood cells successfully detected. Note that the cells are touching each other and boundaries between cells are not distinct. |
