Robust Deep Learning for Computer-Assisted Spinal Surgery

Spinal fusion using pedicle screws is the gold standard technique to treat spinal instabilities. Accurate screw placement is essential due to the proximity of the pedicle to the spinal cord and the related complications in case of co-injury. However, a high level of anatomical knowledge and expertise is required. Different image-assisted techniques enable the visualization of the internal anatomy and facilitate pedicle screw insertion, e.g., fluoroscopy guidance or computed tomography (CT) navigation. This thesis proposes two computer-assisted methods to support the surgeon during both techniques. During fluoroscopy guidance, repeated anatomy-specific standard projections are acquired. Standard projections are X-rays acquired from patient-specific C-arm poses and allow assessing the fracture reduction and implant placement. In the current clinical routine, the C-arm is positioned manually under iterative or continuous fluoroscopy, involving a high radiation dose and time consumption. CT navigation gives an alternative approach for image guidance of spinal interventions. It involves the acquisition of an intraoperative CT in which screw trajectories are manually planned for subsequent 3D navigation. Both image guidance techniques require manual interventions, which are highly expert-dependent, require in-depth knowledge of the anatomy, understanding of anatomical orientation, and increase the procedural time. The methods developed in this thesis should support the surgeon with C-arm positioning during fluoroscopy guidance and in pedicle screw planning during CT-navigation. Many state of the art approaches that propose computer assistance for the manually performed steps make restrictive prior assumptions about modeling or acquisition settings or require external hardware that impedes clinical workflow integration and limits clinical applicability until today. In this thesis, deep learning techniques are employed that learn the anatomical variation from retrospective CT datasets and additional simulations complemented with expert annotations without requiring any other technical equipment