Unsupervised motion artifacts reduction for cone-beam CT via enhanced landmark detection

Motion artifacts in cone-beam computed tomography (CBCT) primarily result from patient movement during the scanning process, which can compromise diagnostic accuracy. Emerging deep learning-based techniques have shown promise in mitigating these artifacts; however, they often rely on motion-free CBCT reconstructions for training, which poses practical challenges in clinical settings. An alternative approach involves leveraging the positions of metallic fiducial markers for motion estimation. While effective, this method is time-intensive and requires additional equipment installation, limiting its practicality. To address these challenges, we propose the Dynamic Landmark Motion Estimation (DLME) method, designed to reduce high-frequency noise and errors in landmark detection, thereby enhancing image quality. DLME is powered by the proposed TriForceNet, a novel landmark detection framework that integrates a sequential hybrid transformer-convolutional neural network architecture, multiresolution heatmap learning, and a multitask learning strategy augmented with an auxiliary segmentation head to improve motion estimation accuracy. Experimental evaluations demonstrate that TriForceNet achieves superior performance compared to state-of-the-art landmark detectors on two-dimensional projection images from the 4D extended cardiac-torso head phantom (XCAT) dataset, real patient CT scans from the CQ500 dataset, and knee regions from the CT scans in the VSD full body dataset. Furthermore, the DLME methodology outperforms traditional unsupervised motion compensation techniques and surpasses supervised, image-based motion artifact reduction methods across these datasets. The source code for the proposed model is publicly available at https://github.com/Thanaporn09/TriForceNet.git.