Ultrasound guidance in navigated liver surgery: toward deep-learning enhanced compensation of deformation and organ motion

Purpose: Accuracy of image-guided liver surgery is challenged by deformation of the liver during the procedure. This study aims at improving navigation accuracy by using intraoperative deep learning segmentation and nonrigid registration of hepatic vasculature from ultrasound (US) images to compensate for changes in liver position and deformation. Methods: This was a single-center prospective study of patients with liver metastases from any origin. Electromagnetic tracking was used to follow US and liver movement. A preoperative 3D model of the liver, including liver lesions, and hepatic and portal vasculature, was registered with the intraoperative organ position. Hepatic vasculature was segmented using a reduced 3D U-Net and registered to preoperative imaging after initial alignment followed by nonrigid registration. Accuracy was assessed as Euclidean distance between the tumor center imaged in the intraoperative US and the registered preoperative image. Results: Median target registration error (TRE) after initial alignment was 11.6 mm in 25 procedures and improved to 6.9 mm after nonrigid registration (p = 0.0076). The number of TREs above 10 mm halved from 16 to 8 after nonrigid registration. In 9 cases, registration was performed twice after failure of the first attempt. The first registration cycle was completed in median 11 min (8:00–18:45 min) and a second in 5 min (2:30–10:20 min). Conclusion: This novel registration workflow using automatic vascular detection and nonrigid registration allows to accurately localize liver lesions. Further automation in the workflow is required in initial alignment and classification accuracy.