In volume visualization, the definition of the regions of interest is inherently an iterative trial-and-error process finding out the best parameters to classify and render the final image. Generally, the user requires a lot of expertise to analyze and edit these parameters through multi-dimensional transfer functions. In this thesis, we present a framework of methods to label on-demand multiple regions of interest. The methods selected are a combination of 1vs1 Adaboost binary classifiers and an ECOC framework to combine binary results to generate a multi-class result. On a first step, Adaboost is used to train a set of 1vs1 binary classifiers, with a labeled subset of points on the target volume. On a second step, an ECOC framework is used to combine the Adaboost classifiers and classify the rest of the volume, assigning a label to each point among multiple possible labels. The labels have to be introduced by an expert on the target volume, and this labels have to be a small subset of all the points on the volume we want to classify. That way, we require a small e↵ort to the expert. But this requires an interactive process where the classification results are obtained in real or near real-time. That why on this master thesis we implemented the classification step in OpenCL, to exploit the parallelism in modern GPU. We provide experimental results for both accuracy on classification and execution time speedup, comparing GPU to single and multi-core CPU. Along with this work we will present some work derived from the use of OpenCL for the experiments, that we shared in OpenSource through Google code, and some abstraction on the parallelization process for any algorithm. Also, we will comment on future work and present some conclusions as the final sections of this document.