This work demonstrates the development and application of fast deep-learning models designed for the mitigation of noise in Monte Carlo dose distributions (MC-DDs) with high statistical uncertainty (SU) of radiotherapy treatment plans. Five 3D U-net models were developed, varying in input, input size and batch normalization. The models were trained on pairs of high/low SU MC-DD pairs of randomly generated clinically motivated treatment plans calculated on open source available computed tomography (CT) scans. Depending on the model, the CT was included as input. The accuracy of the models is evaluated on the test set using gamma passing rate (2% global, 2 mm, 10% threshold) comparing denoised and low SU MC-DD. The best performing model included the CT input and no batch normalization. It was retrained for detailed evaluation on inhouse data, consisting of high/low-SU MC-DD pairs generated from 106 clinically-motivated volumetric modulated arc therapy (VMAT) arcs for 29 CT scans. The dataset was augmented to encompass a total of 3074 pairs. On the test set, the denoised MC-DDs agree with low-SU MC-DDs, by an average (standard deviation) gamma passing rate of 82.9% (4.7%). Applied to 12 previously unobserved clinically-motivated cases originating from different treatment sites yields an average gamma passing rate of 91.0%. Denoised DDs are computed, on average, in 35.1 seconds, a 340-fold efficiency gain compared to the calculation of low-SU MC- DDs
