Cone-beam computed tomography (CBCT) has been widely used in various fields such as industrial measurement, security inspection, and medical imaging. However, during the CT scanning process, the Compton effect inevitably leads to issues in CT images, including scatter artifacts, reduced image contrast, and information loss. Compared with other types of scanning methods, CBCT has a larger cone angle and detector area, which allows it to receive more scattered photons, thus being more severely affected by scatter. As a result, effective methods need to be adopted to perform scatter correction on CBCT images, restore image quality, and improve the accuracy of clinical diagnosis.

To achieve CBCT scatter correction, we design a novel dual-encoder network model, DEU-Net. DEU-Net is composed of a densely connected convolutional neural network (DCCNN) module and a Swin transformer (Swin-T) module. These two modules are utilized to extract local and global features of the image and combine them to achieve preliminary image correction in the image domain. Based on the low-frequency characteristics of scatter, the DEU-Net model is combined with a two-dimensional discrete wavelet transform. Scatter information is extracted in the projection domain to achieve final scatter correction. A weighted loss function is designed to ensure that the model training process focuses more on the parts of the image with complex structures and large errors, thereby obtaining better correction results.

The feasibility and effectiveness of this method are verified using the MC dataset, which is composed of pelvic data and high- and low-dose data. The comparison results before and after correction are shown in Fig. 5. The proposed method is compared with other deep learning-based methods, and the results are presented in Fig. 6, with the indicator data shown in Table 1. It is evident from the results that, compared with other correction methods, the method proposed in this paper has the smallest CT difference and the best correction effect. The corrected CT images exhibit higher contrast, which makes the distinction between different tissue structures more obvious. Moreover, ablation experiments are conducted to verify the positive effects of each module on the correction results (Fig. 7 and Table 2). The proposed method can also achieve excellent results in correcting scatter in low-dose CT images (Figs. 10 and 11, Tables 3 and 4). This demonstrates its potential clinical values in restoring image quality and data accuracy and realizing scatter correction. In addition, this method is used to correct and analyze the CT images of turbine blades to verify its ability to correct real scatter artifacts. The results are presented in Fig. 13, as well as Tables 5 and 6. The CT images corrected by the method in this paper are cleaner, the scatter artifacts are better suppressed, and the grayscale distribution of the object part is more uniform.

During the CBCT and imaging process, the Compton effect affects the quality of the reconstructed image, thus leading to phenomena such as scatter artifacts and inaccurate CT values. We propose a novel DEU-Net structure to achieve preliminary scatter correction in the image domain. Moreover, based on the low-frequency characteristics of scatter, the output of the model is combined with two-dimensional discrete wavelet transform to extract the low-frequency scatter signal in the projection domain, thereby realizing the final scatter correction of CBCT images. In this network structure, the two encoding paths serve different purposes. DCCNN module is used to extract local features of the image, while Swin-T module is used to extract global features. These two modules complement each other, thereby enhancing the model’s feature extraction ability and improving the correction effect. In addition, a new weighted loss function is designed to ensure that the model training process gives more attention to the parts of the object with complex structures. The experimental results show that the method combining the dual-encoder network model with wavelet transform can effectively perform scatter correction, improve the quality of CBCT images, and has the potential to enhance the accuracy of radiotherapy diagnosis in clinical practice. Meanwhile, applying this method to perform scatter correction on the CT images of aero-engine turbine blades verifies its ability to correct real scatter artifacts.