Trainable grey-level models for disentangling overlapping chromosomes

We propose and evaluate a mechanism for resolving the segmentation of overlapping chromosomes using trainable models of the expected banding appearance. The models consist of templates of sub-chromosome length band pro"les. Candidate chromosome segments are classi"ed according to their responses to the entire set of templates, and matched on the basis of the classi"cations. Evaluation of the models using a set of annotated banding pro"les yields correct classi"cation rates of 90.8% for isolated chromosomes, and 55.4% for chromosome fragments; 70.6% of overlapping chromosome pairs, simulated using the pro"le data set, are correctly resolved. ( 1999 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.