Abstract
2D X-ray images are extensively employed for intraoperative navigation and localization owing to their high imaging efficiency, low radiation risk, and affordability. However, this method can only yield overlapped anatomical information from a restricted number of projected views. Conversely, intraoperative CT scanning techniques, offering 3D images, elevate the risk of radiation exposure for both patients and healthcare professionals. For this purpose, we propose a V-shaped convolutional attention mechanism network (X-CTCANet) designed for X-ray reconstruction of CT images. The network enhances reconstruction performance by promoting task consistency in encoding-decoding, minimizing semantic differences between feature mappings. Additionally, it introduces an adaptive convolutional channel attention (CCA) mechanism to compel the network to prioritize essential feature regions. Experimental results demonstrate the successful CT image reconstruction from spine X-rays using X-CTCANet, achieving an SSIM value of 0.805 and a PSNR value of 34.64 dB. This underscores the considerable potential of accurate 3D CT reconstruction from 2D X-ray images in offering image support for surgical robots.