Recently we have witnessed the impact of deep neural networks (DNN) to the field of computational imaging (CI). Usually, a DNN should be trained on a specific training dataset before it can be used to solve the corresponding CI problem. The DNN trained in this way is fundamentally a black box, lack of deep understanding of its applicability. Here we report our recent works on physics-driven untrained deep neural networks (PhysenNet) for computational imaging. The PhysenNet employs a strategy that incorporates a physical imaging model into a conventional DNN. PhysenNet has two apparent advantages. First, it does not need to be trained on any dataset. Instead, it just requires the data to be process as its input. The interplay between the physical model and the randomly initialized DNN provides a mechanism to optimize the DNN, and produce a good reconstruction. Second, the reconstructed image satisfies the constraint imposed by the physical model so that it is interpretable. We will take phase imaging and computational ghost imaging as examples to demonstrate the principle.