3D ultrasound shape completion and anatomical feature detection for minimally invasive spine surgery.

Ultrasound (US) with 3D reconstruction is being explored to offer a radiation-free approach to visualizing anatomical structures. Such a method could be useful for navigating and assisting minimally invasive spine surgery where direct sight on the surgical site is absent. During surgery, the pre-operative CT model and surgical plans are registered to the patient's anatomy by using intra-operative US reconstruction. However, accurate and automatic registration remains challenging. This difficulty arises from an incomplete detection of the bone geometry in US images and the challenges in identifying anatomical landmarks. To address the problem, this work presents a pipeline to automate the workflow by offering an initial CT-to-US registration. This work utilizes PointAttN for 3D shape completion that completes occluded bone structures from partial US reconstruction. This enriched point cloud is then segmented using PointNet++ to identify specific anatomical features. To train the network, synthetic 3D representations of partial views are generated from fifty CT models of the lumbar spine by simulating US physics, effectively mimicking the intraoperative scenario. The proposed work yields a mean registration error of 1.34 mm and 1.63 mm on real US reconstructions of agar phantoms and an ex vivo human spine, respectively. This comprehensive 3D representation enhances anatomical feature interpretation, enabling robust automatic registration. The clinical potential of this framework merits further investigation in pre-clinical trials.