Collections of classifiers tuned for cell finding with an application to building digital cell atlases of drosophila embryos

The different combinations of genes active in different cells control the development and diversity of multicel-lular organisms. Yet the codes that control this process, written in both cis–regulatory and protein–coding DNA sequence, are poorly understood. This is despite the availability of imaging techniques that allow detection of gene activity at the resolution of single transcription units on chromosomes in individual interphase diploid nuclei[1]. While it has been shown that digital cell atlases are indeed possible[2], producing query-able models from imaging for Systems Biology[3] is far from commonplace for even the simplest of organisms. While much algorithmic work has been done on transforming pixels to meaningful objects (i.e. segmenting cell nuclei)[4], this necessary first step is far from a solved problem. Algorithm tuning remains a critical component and requires expertise in computer vision and image processing. This expertise is called upon repeatedly with new experimental data: when changes in the model organism, hybridization method, or the microscope result in changes to the generated images. We have successfully developed adaptive segmentation methods[5] that, while tuned for particular types of images, are tuned using machine learning methods. The advantage machine learning provides is rather than have the experimentalist tune the computer vision parameter(s) directly—this, again, requires an understanding of computer vision—the experimentalist instead tunes the system by providing direct feedback. The feedback mechanism is designed to be a of a natural form: given a collection of segmented pixels, the experimen-talist labels each collection as " a partial nucleus " , " a combination of two nuclei " , " a complete nucleus " , etc. From these these labeled collections we build classifiers that mimic the experimentalist's recognition of various segments. Using the human labeled examples as input, classifiers are built using JBoost[6], an open source, Java implementation of the Adaboost machine learning algorithm[7]. Adaboost generates a mapping from segments to scores, where high scores correspond to high confidence that the segment is correct (Figure 1a–g). Adaboost is a general-purpose learning algorithm which is particularly powerful when combining many weakly predictive features. The predictive features in this case are the values of the morphological features of the segments (Figure 1f). JBoost generates classifiers in a form of decision tree called Alternating Decision Trees (ADT)[8]. The ADT classifier (Figure 1g) acts a as proxy for the experimentalist, and we use these classifiers to guide the search for proper segmentations. As the experimental conditions change …