Bisque: A Scalable Biological Image Database and Analysis Framework - High-Throughput Image Reconstruction and Analysis - page 300

interpretation of the image and gives no measure of the underlying data uncer-
tainty. In an alternative approach, each pixel is classified as part of the cell or
part of the background. More formally, let S be an image with N pixels and X =

X s , S be the set of corresponding random variables, where X s is 0 or 1 depend-

ing on whether it is part of the background or cell. A possible world is an instan-
tiation of all random variables, and thus defines a specific binary image. If images
are 1,024

s

∈

1,024, then we have 2 1,024 × 1,024 possible worlds. Exploring this entire
space is clearly intractable, and we need a method to overcome this limitation.
One possible solution to the computational bottleneck is by sampling over
the possible worlds. A simple approach is to assume pixel independence. In this
method, we randomly classify each cell as part of the cell according to its probabil-
ity. We repeat this operation for all pixels to obtain one possible world. Features
can then be measured in this binary world. This process is then repeated until a
sufficient sample size of worlds has been collected. A distribution of the features is
obtained in this manner. A better and less error-prone approach is to assume that
pixels are inherently dependent upon their surrounding neighbors. One can de-
fine a neighborhood system and assume that pixels are conditionally independent
given their neighborhood [18]. The challenge then is to estimate the joint proba-
bility distribution. For this, one can employ Gibbs sampling, a common Markov
Chain Monte Carlo (MCMC) technique [7]. In preliminary experimental results,
this technique gave good results for a number of cellular features when compared
with manual ground truth.

×

13.5 Concluding Remarks

Data-driven science has become a reality, and quantitative image analysis is becom-
ing ever more important to many scientific endeavors. This chapter discusses new
developments in data-driven image-based science, bridging the gap between image
analysis and databases. Bisque extends traditional image databases with an ex-
tensible analysis integration framework. It is built using a distributed Web-based
architecture to ensure scalability and ease of access for both users and analy-
sis designers. The combination of a flexible schema architecture with ontologies
(controlled vocabularies) promises easier data modeling for experts in many fields
without the need to consult with database experts. Managing provenance of im-
ages and chains of analyses improves the development of new hypotheses and the
reproducibility of conclusions. New techniques in managing uncertainty and prob-
abilistic data have been explored, capturing nuances in data previously unseen.
Bisque is a combination of emerging techniques and technologies providing an in-
tegrated package for scientists, image processing experts, and database researchers
in order to explore this new domain.

References

[1] Proceedings of the 23nd International Conference on Data Engineering, ICDE 2007,

April 15--20, 2007, Istanbul, Turkey. IEEE, 2007.

[2] C. Aggarwal. ''On density based transforms for uncertain data mining.'' In International

Conference on Data Engineering , 2007.