Conventional ICA employs the assumption that data can be decomposed into statistically independent sources and typically models the probability density functions of the underlying sources as symmetric and highly kurtotic. However, when source data violate these assumptions, conventional ICA methods might not work well. As a result, modeling of the underlying sources becomes an important issue for ICA applications. In this paper, we propose a source density-driven adaptive ICA (SDA-ICA) method for functional magnetic resonant imaging (fMRI) which allows the utilization of a priori information. The SDA-ICA algorithm involves a two-step procedure. It uses a conventional ICA algorithm to obtain initial independent source estimates at the first-step; then source density calculation using a Kernel estimator technique, and finally a refitted "optimal" nonlinearity based source density function is obtained for each source separation at the second step. Since the physiological meaningful components (e.g., activated regions) of fMRI signals are typically governed by small percentage of the whole brain map on a task-related activation, extra prior information (using a skewed-weighted distribution transformation) is applied to the algorithm on the regions of interest of data for emphasizing the importance of the tail part of the distribution. Our experimental results show that the SDA-ICA method can provide flexible source ada privity and improve ICA performance on real fMRI data compared with other methods.