In this paper we define the 3d topological map and give an optimal algorithm which computes it from a segmented image. This data structure encodes totally all the information given by the segmentation. More, it allows to continue segmentation either algorithmically or interactively. We propose an original approach which uses several levels of maps. This allows us to propose a reasonable and implementable solution where other approaches don't allow suitable solutions. Moreover our solution has been implemented and the theoretical results translate very well in practical applications.