Additive manufacturing (AM) is a layer-by-layer manufacturing method which is growing considerably in recent years. Inner dimensional measurement based on Computed Tomography (CT) image is an irreplaceable method for quality control in AM. However, for special AM components, limited-angle CT imaging is the only method because of physical limitations. CT reconstruction image with ghost artifacts through incomplete data can’t be used to implement dimensional measurement. To deal with this problem, a method based on encoder-decoder neural network is proposed and applied to characterize the dimensional parameter directly from CT sinogram without reconstruction. Experiment results showed that when characterizing AM component with length of side from 10 to 15mm, the relative accuracy is less than 18.26% and the relative accuracy is less than 10% in over 83.5% sample, which verifies the effectiveness of the method that can be applied in dimensional measurement.