Learning-based dual-domain rigid motion estimation in interventional C-arm cone-beam CT

We introduce a novel approach for rigid patient motion estimation in interventional C-arm cone beam computed tomography for acute stroke patients. Extended acquisition times compared to diagnostic CT increase the likelihood of patient motion, leading to streaking and blurring artifacts in the reconstructed images. Several state-of-the-art motion compensation strategies exist for different domains of the imaging process, including reconstruction-free data consistency assessments, using epipolar consistency conditions, and reconstruction-based autofocus methods, optimizing metrics like gray-value histogram entropy or total variation. Consistency-based approaches are merely sensitive to out-of-plane motion, while autofocus methods can detect in-plane motion but tend to converge into local minima for larger motion amplitudes. We aim to leverage strengths and mitigate limitations of both approaches by combining them in a deep learning-based model. Our method consists of two parallel 50-layer residual networks followed by a classification network, trained and tested using 5-fold cross-validation on 10,500 simulated motion trajectories applied to 21 motion-free clinical acquisitions. The network predicts 6 rigid directional displacement probabilities for each of the 496 projections, yielding a 496x6-dimensional classification problem. First experiments indicate good classification performance, particularly in discriminating between in- and out-of-plane motion, with an average ROC-AUC of 0.9527 and PR-AUC of 0.3961. This encourages further investigations and model fine-tuning, which is subject to ongoing research.