In many vision problems, the performance of the segmentation step is highly dependent on the algorithm selection and its parametrization. These tasks are tricky and time-consumming. In this paper, we present an approach to perform task-oriented segmentation based on segmentation algorithm parameter tuning and learning techniques. We propose a scheme that, for each segmentation algorithm to test, first extracts optimal parameters and second learns the region labeling according to the segmentation task. This supervised approach uses two kinds of groundtruth data: manual region-based segmentations and semantic region labels. The first step consists in extracting optimal segmentation algorithm parameters by using a closed-loop optimization procedure, an evaluation metric and groundtruth (manual region segmentations). During the second step, region classifiers are trained based on ground-truth annotations (semantic region labels) to allow segmentation labeling. This knowledge (i.e. optimal parameters and learned region classifiers) is then used to produce an optimized class-based segmentation of new images. Our main contribution is to propose a methodology to easily set up the segmentation task in vision systems. Our method only requires the user to provide segmentation algorithms and labeled ground-truths. The experiment on a biological application shows that our optimization scheme is reliable for different state-of-the-art segmentation algorithms. Evaluation of results also demonstrates that the optimized classbased segmentation achieves a better level of accuracy than the non-optimized class-based segmentation approach.