Flexible synapse detection in fluorescence micrographs by modeling human expert grading

A particularly difficult task in molecular imaging is the analysis of fluorescence microscopy images of neural tissue, as they usually exhibit a high density of objects with diffuse signals. To automate synapse detection in such images, one has to simulate aspects of human pattern recognition skills to account for low signal-to-noise-ratios. We propose a machine learning based method that allows a direct integration of the experts' visual expertise who tag a low number of referential synapses according to their degree of synapse likeness. The sensitivity and positive predictive values show that by using graded likeness information in our learning algorithm we can provide an intuitively tunable tool for neural tissue slide evaluation.