אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Fluorescent microscopy imaging has become one of the main tools of biological research. Most wet labs collect Gigabytes of images every day. The resulting data is almost exclusively annotated and analyzed by hand. Such data contains a wealth of information potentially relevant to a variety of research questions. Yet, due to the labor-intensive nature of data analysis and huge volume, the precious data is discarded without a chance to realize its full scientific value. A typical setting contains several cells imaged together concurrently on several channels. In this flat world of simple morphology, there is surprisingly little automation of image analysis and data processing. In this talk I will discuss the potential of machine learning methods to revolutionize the field and accelerate bio-medical discovery by changing the way researchers are collecting, analyzing and sharing vast amounts of data in collaborative distributed environments. I will present two case studies. The first is automated classification of metaphase mitotic cells based on phenotypic differences. We automatically segment individual cells and detect phenotypic features in raw microscope images. Using machine learning methods, we perform three-way classification of the mitotic cells into monopolar, bipolar and multipolar in order to do high throughput gene analysis. In the second we use insight from natural scene segmentation algorithms to automatically segment live human cells solely from phase images, in the interest of leaving fluorescent channels to track arbitrary unrelated signals. We apply the approach of texture decomposition and classification in this 2D rotation-invariant universe and compare the performance to the natural 3D world picture decomposition. In conclusion we consider the issue of method generalization--a reproducible computational protocol designed to augment a "wet lab" protocol in order to handle a wide range of cell lines and organisms.

Host: Itai Yanai