Fast and accurate scatter correction is essential for quantitative positron emission tomography (PET). However, single scatter simulation usually disregards multiple scattering as well as scatter from outside the field of view and relies on error-prone tail fitting for scatter scaling. The use of Monte-Carlo scatter simulation, a highly accurate alternative, is impeded by long computation times. Deep scatter estimation (DSE) uses convolutional neural networks (CNN) which are trained to reproduce known scatter distributions from known input data, and later applied to unknown input data for fast scatter estimation with similar accuracy. This work investigates DSE for PET imaging. Methods: We collected data from 20 patients and simulated reference scatter distributions for each of 71 bed positions. We implemented a CNN which inputs measured PET emission data as well as attenuation correction factors, and trained the network to reproduce the scatter distributions of 57 (training) bed positions. We then used the model to predict scatter estimates for the 14 remaining (validation) bed positions, and reconstructed PET images using reference and estimated scatter distributions for scatter correction, respectively. Normalized mean absolute errors (NMAEs) were computed for scatter estimates and reconstructed PET images. All computations were run on a workstation with a graphics processing unit. Results: Model training took 32 h for 20 epochs, resulting in DSE durations of 5.3 s per bed position. Scatter NMAEs for 13 of 14 bed positions ranged from 4 to 10 % (mean, 7.1 %), with one outlier (14 %) in a bed position where the filled bladder extended outside the axial FOV. In reconstructed PET images, these NMAEs translated to 1 to 8 % (mean, 3.6 %; outlier: 28 %). Conclusion: Convolutional neural networks have the potential to significantly speed up scatter estimation, opening the door to faster and more accurate scatter correction in clinical PET.