Automatic signal classification in fluorescence in situ hybridization images

Background: Previous systems for dot (signal) counting in fluorescence in s itu hybridization (FISH) images have relied on an auto-focusing method for obtaining a clearly defined image. Because signals are distributed in three dimensions within the nucleus and artifacts such as debris and background fluorescence can attract the focusing method. valid signals can be left unf ocused or unseen. This leads to dot counting errors, which increase with th e number of probes. Methods: The approach described here dispenses with auto-focusing, and inst ead relies on a neural network (NN) classifier that discriminates between i n and out-of-focus images taken at different focal planes of the same field of view. Discrimination is performed by the NN, which classifies signals o f each image as valid data or artifacts (due to out of focusing). The image that contains no artifacts is the in-focus image selected for dot count pr oportion estimation. in fluorescence in situ Results: Using an NN classifier and a set of features to represent signals improves upon previous discrimination schemes that are based on nonadaptabl e decision boundaries and single-feature signal representation. Moreover, t he classifier is not limited by the number of probes. Three classification strategies, two of them hierarchical, have been examined and found to achie ve each between 83% and 87% accuracy on unseen data. Screening, while per f orming dot counting, of in and out-of-focus images based on signal classifi cation suggests an accurate and efficient alternative to that obtained usin g an auto-focusing mechanism. (C) 2001 Wiley-Liss. Inc.