Cone-beam tomography enables rapid 3D acquisitions, making it a suitable imaging modality for dental imaging. However, as with all X-ray techniques, the main challenge is to reduce the dose while maintaining good image quality. Moreover, dental reconstructions face a series of issues stemming from truncated projections as well as metal and cone beam artifacts. The aim here is to investigate the ability of neural networks to improve the quality of 3D CBCT dental images at low doses. We test different configurations of convolutional neural networks, trained in a supervised way to reduce artifacts and noise present in analytically reconstructed volumes. In a study on 32 experimental cone beam volumes, we show their capacity to preserve and enhance details while still reducing the artifacts. The best results are obtained with a 3D U-Net which compares advantageously with a TV regularized iterative method and is considerably faster.