Compared to computed tomography (CT), three-dimensional rotational angiography (3DRA) presents tomographic images with relatively higher noise due to a lack of imaging projections and a more considerable scatter volume. This paper extends the use of the residual encoder-decoder convolutional neural network (RED-CNN) to denoise 3DRA images with a specific mode of training, where an in-house algorithm was trained using CT images of homogeneity water phantom as target noise level for 3DRA images to refer to. The algorithm was applied to 3DRA images obtained using GE Optima CL323i with CT images from a Siemens Somatom 16 as benchmarks. All 3DRA images were also denoised using the standard block-matching and 3D filtering (BM3D) method as a preliminary comparison. Noise levels were quantified in Hounsfield unit using ImQuest software, while the limiting spatial resolution (10% modulation transfer function, MTF) was determined using Spice-CT plugin of the ImageJ tool. Whereas resulting 3DRA images denoised using BM3D had an average of 22,6% decrease in noise level (from 29,6 HU to 22,9 HU), our method presents an average of 49,0% decrease in noise level (29,6 HU to 15,1 HU) for the same images. The 3DRA images denoised using RED-CNN underwent an increase of spatial resolution (8,2%), while BM3D-denoised images had their 10% MTF reduced by an average of 8,7%. These preliminary results had indicated a potential use of RED-CNN based method to denoise 3DRA images without compromising the spatial resolution.