A series of multi-contrast MR images are usually required in various MR applications, such as T1 and T2 mappings, which provide quantitative information of inherent tissue properties for diagnosis purpose. However, its clinical application is limited by the long scanning time. The emerging theory of compressed sensing has shown great potential in accelerating MR acquisitions. Recently, a principal component analysis based method has been proposed exploiting the temporal sparsity via truncated PCs based on the knowledge of the analytic model and possible parameter range. However, it may generate model errors when such prior information is not accurate. In this work, the support of the PC coefficients were detected in a more adaptive way using iterative support detection. Reconstructions based on two knee data sets were conducted to demonstrate the effectiveness of the proposed method.