Region Growing in Nonpictorial Data for Organ-Specific Toxicity Prediction

Region growing is a well-known concept in image processing that, among other things, effectively contributes to mining of pictorial data. The goal of this research effort is to (1) investigate region growing in nonpictorial data and (2) determine the effectiveness of the regions in mining of such data. Part (1) is met by introducing a new version of the self-organizing map (SOM), Neighborly SOM, capable of delivering such regions. Part (2) is met by introducing a new prediction methodology using the delivered regions and measuring its effectiveness by (a) applying the method to 10 pairs of training and test sets [repeated random sub-sampling (RRSS) cross-validation] predicting the chemical agents' liver toxicity and (b) comparing the liver toxicity prediction accuracy with the predictions produced by C4.5, and the traditional SOM using leave-one-out (LOO) and RRSS cross-validations. The results revealed that the proposed methodology has a better performance.