3D synthetic CT patch generation and reconstruction by using multi-resolution generative adversarial network

Cone-beam CT (CBCT) offers several advantages, such as lower radiation doses, improved image quality, more accurate image reconstruction, and faster image display, essential for image-guided radiation therapy and supporting patient diagnosis and treatment. Learning-based methods can significantly enhance CBCT image quality and HU accuracy, aiding in dosimetry calculations. In this study, we introduced a new CBCT-to-tensor mapping range to improve the activation function's computational performance. Additionally, a new deep separable feature extraction module was incorporated into the 3D Double-Chain-CycleGAN (DCC-GAN), and patch-based training and inference methods were employed to reduce hardware costs. Experimental results show that the 3D DCC-GAN model outperforms its 2D-based counterpart, with a 6.67% improvement in mean absolute error (MAE), a 5.14% improvement in root mean squared error, a 0.27 increase in peak signal-to-noise ratio from 27.80 to 28.07, and a 6.4% improvement in Gradient Magnitude Similarity Deviation. Furthermore, the 3D DCC-GAN model requires significantly less GPU memory than similar models, even those based on 2D CBCT images.