In this talk, I will show that several mathematical models in imaging sciences, such as the sparsity-based models and statistical models, can be reformulated as deep learning models. The basic idea is that the iterative optimization algorithms for energy minimization or statistical inference can be unfolded to be deep architectures. In this way, the parameters and even the formulations of these models can be discriminatively learned for specific task.I will show that the Markov random field model in image prior modeling, iterative shrinkage in signal processing, compressive sensing model in MRI can be formulated to be deep learning problems. These induced deep architectures are non-conventional, task-specific and achieved state-of-the-art results for solving image inverse problems, e.g., image restoration, compressive sensing MRI.