Thanks to the recent success of generative adversarial network (GAN) for image synthesis, there are many exciting GAN approaches that successfully synthesize MR image contrast from other images with different contrasts. These approaches are potentially important for image imputation problems, where complete set of data is often difficult to obtain and image synthesis is one of the key solutions for handling the missing data problem. Unfortunately, the lack of the scalability of the existing GAN-based image translation approaches poses a fundamental challenge to understand the nature of the MR contrast imputation problem: which contrast does matter? Here, we present a systematic approach using Collaborative Generative Adversarial Networks (CollaGAN), which enable the learning of the joint image manifold of multiple MR contrasts to investigate which contrasts are essential. Our experimental results showed that the exogenous contrast from contrast agents is not replaceable, but other endogenous contrast such as T1, T2, etc can be synthesized from other contrast. These findings may give important guidance to the acquisition protocol design for MR in real clinical environment.