We propose a deep learning approach to visualize vessels during image-guided procedures without requiring contrast agents or preoperative 4D CT scans.
X-ray images are routinely acquired during image-guided surgical interventions, but vessels are not visible in these images.
However, due to their nephrotoxicity, contrast agents may only be used in limited quantities during procedures. As an alternative, a preoperative vessel segmentation can be projected onto the images.
Because anatomy deforms during the intervention, the vessel segmentation must be updated. Using a neural network, we can predict anatomical deformation in X-ray images to accurately superimpose vessels in their correct position and shape. Our comparison shows uncorrected hepatic veins (red) that fail to track breathing motion and deviate from their true position. When updated via the neural network prediction (green), the vessels accurately follow the breathing motion visible in the image.
