Preparing a decent word reference is the way in to an effective picture combination technique for scanty portrayal based models. In this paper, we propose a clever word reference learning plan for clinical picture combination. To begin with, we build up the feeble data of pictures by removing and adding their multi-facet subtleties to produce the educational patches. In the meantime, we present a straightforward and compelling multi-scale testing to execute a multi-scale portrayal of patches while lessening the computational expense. Second, we plan a local energy metric and a multi-scale spatial recurrence metric for bunching the picture patches with a comparable splendor and detail data into each particular fix bunch.