This paper introduces a new approximation method called Deep Multi Spline (DM-Spline) which is based on a deep learning model and various types of spline functions (B-Spline, UAT-Spline, hyperbolic-Spline, ...). The main idea is to combine different spline functions into a learning model that controls their contribution to the accurate approximation of the function to be built. We introduce a new bridge between Deep Neural Networks (DNNs) and spline approximation methods by developing a detailed theory in each layer and presenting some results on this subject. To test the robustness of the proposed approach, a comparison to state-of-the-art is achieved and shows the efficiency of our model. An application in the medical field is also provided.