Dual-domain attention-guided convolutional neural network for low-dose cone-beam computed tomography reconstruction

Excessive ionizing radiation in cone-beam computed tomography (CBCT) causes damage to patients, whereas a low radiation dose degrades the imaging quality. To improve the quality of low-dose CBCT images, deep-learning-based methods are developed and have obtained good performance. However, most previous studies only process the reconstructed CT images, and cannot recover the structures already lost in the reconstruction process. In this paper, a dual-domain attention-guided network framework (Dual-AGNet) is developed to process images in both projection and reconstruction domains. Spatial attention modules are included in the AGNet to effectively and adaptively compensate the intra- and inter-images information in both domains. Moreover, a joint loss function is developed to circumvent the structures loss and over-smoothness in CT images. Our method is evaluated and compared with the state-of-the-art methods on a simulated and a real low-dose CBCT datasets of walnuts. Our Dual-AGNet obtains significantly better performance than the state-of-the-art methods; on the simulated and real datasets, it decreases the root mean square error by approximately 11% and 19%, increases the peak signal-to-noise ratio by approximately 5% and 7%, and increases the structural similarity by approximately 5% and 2%, respectively. In qualitative evaluation, our Dual-AGNet not only suppresses the noise, but also provides realistic CT images with many delicate structures. In addition, the developed Dual-AGNet can be integrated into the existing CBCT system to promote the development of low-dose CBCT imaging. Testing code is available at https://github.com/LianyingChao/Dual-AGNet . • A novel dual-domain deep learning framework for low-dose CT reconstruction. • A 3D spatial attention module for well utilizing the intra- and inter-images information. • A novel joint loss function for circumventing the structures loss and over-smoothness. • Consistently good performance was obtained on both simulated and real datasets.