Cone-beam computed tomography (CBCT) enables three-dimensional imaging of the internal structure of objects in a non-invasive way with high accuracy. Practical misalignment of the CBCT system causes geometric artefacts in reconstructed images, which seriously degrades image quality in ways such as detail loss and decreased spatial resolution. This leads to inaccurate distinction of defects in detection, especially in precise industrial fields like aerospace and instrument manufacturing. This paper presents a method to reduce the geometric artefacts based on a data-driven strategy, which is an end-to-end modified fully convolutional neural network (M-FCNN). The designed M-FCCN contains five convolution layers and five deconvolution layers for feature extraction and output image rebuilding, respectively. In addition, the pooling layer is not used in the designed M-FCNN, considering the preservation of details in the reconstructed image. In this M-FCCN, artefact images with different features have been trained separately. After training, the M-FCNN can be applied to directly reduce geometric artefacts in the reconstructed image. The designed M-FCNN has been demonstrated with different types of synthetic data and has achieved accurate results. It is also validated with practical data, including carbon composite and medical oral phantoms with comparable quality to phantom-based methods, proving that it is an effective way to reduce geometric artefacts in the image domain by means of a data-driven strategy.