In general, there are two open challenges for domain specific visual concept detection. First is the high intra-class variations and second is to collect large collection of sample training data covering the wide variety. In this research we present a novel medical image sampling approach to handle these two challenges. For huge intra-class variations present in the data collection we propose a unique clustering method to group similar data samples. We propose to measure similarity on the basis of membership degree between candidate groups. At the same time we handled the issue of large data collection over wide variety. We synthetically re-sampled the data on the basis of membership score. The membership score helped to make use of minor groups which has small sample size. Experiments show that our proposed method can achieve promising results and outperforms existing approaches particularly for medical image concept detection. © 2019 IEEE.