Acoustic simulations used in the articulatory synthesis of speech take a series of vocal tract shapes as an input. Acoustic simulations assume a plane wave propagation, simplifying and limiting the calculation time. It is, therefore, necessary to split 2D vocal tract shapes into small tubes perpendicular to the centerline simulating the plane wave propagation. The algorithm developed previously used a time-consuming regularization step whose computation time was close to that of acoustic simulations. Therefore, we explored the possibility of using deep learning to perform this step and accelerate the whole synthesis process. We used a database with a large number of rt-MRI images (150 000) and our regularizing algorithm for training. Two architectures were tested, one using a regression strategy applied to the two curves defining the vocal tract and one exploiting the classification of pixels in 2D images of the vocal tract. The first turned out to be much faster, even if it requires checking that the center line is correct and, in some sporadic cases using the initial algorithm as a fallback solution.