In this paper, we discuss the use of graph-cuts to merge the re gions of the watershed transform optimally. Watershed is a simple, intuitive and efficient way of segmen ting an image. Unfortunately it presents a few limitations such as over-segmentation and poor detection o f low boundaries. Our segmentation process merges regions of the watershed over-segmentation by minimizing a specific criterion using graph-cuts optimization. Two methods will be introduced in this paper. The first is bas ed on regions histogram and dissimilarity measures between adjacent regions. The second method deals with efficient approximation of minimal surfaces and geodesics. Experimental results show that the se techniques can efficiently be used for large images segmentation when a pre-computed low level segmenta tion is available. We will present these methods in the context of interactive medical image segmentation.