We investigate the fuzzy ARTMAP (FA) in off and online image classification for diagnosis of genetic abnormalities. We evaluate the classification task (detecting abnormalities separately or simultaneously), classifier paradigm (monolithic or hierarchical), ordering strategy (averaging or voting), training mode (for one epoch, with validation or until completion) and sensitivity to parameters. We find the FA accurate in achieving the tasks requiring only few training epochs. Superiority is found for the voting strategy and training until completion mode. Compared to other classifiers, the FA does not loose but gain accuracy when overtrained. Its accuracy is comparable with those of the multi-layer perceptron and support vector machine and superior to those of the naive Bayesian and linear classifiers.