AI-Based Surgical Planning Systems

Abstract

Before operating, a surgeon studies the patient's CT scan and tries to build a mental picture of what lies beneath the skin-- where the tumour ends, where the artery begins, how the bile ducts branch, whether there is enough healthy liver left to survive the resection. This mental reconstruction happens in every operating room, every day, and most of the time it works well enough. But when the anatomy is complex -- a liver tumour wrapped around the portal vein, a brain tumour abutting the motor cortex, a shattered pelvis with fifteen bone fragments -- the surgeon's spatial imagination reaches its limits, and planning errors creep in. By one estimate, anatomical misinterpretation contributes to 18 to 28 percent of surgical complications in hepatobiliary and neurosurgical procedures. AI-based surgical planning systems aim to fix this by automating the steps that a human planner does slowly and imperfectly: segmenting the organs and lesions from the scan, building a three-dimensional model the surgeon can rotate and examine from any angle, optimising the surgical approach to avoid critical structures, and predicting complications before the first incision is made. We evaluated 180 such systems across five surgical specialties -- orthopaedic (n = 48), cardiac (n = 36), hepatobiliary (n = 32), head-and-neck (n = 28), and neurosurgery (n = 36) -- tested in our laboratories in Tallinn, Madrid, and Zurich between 2019 and 2023. We scored each system on segmentation accuracy, 3D model quality, how much it shortened planning time, how well it predicted complications, and whether surgeons actually liked using it. The resulting composite score - the Surgical Planning AI Quality Index, or SPAQI -- correlated with clinically significant operative improvement at r = +0.84 (AUC = 0.884). In a prospective 96-case comparison, AI-planned operations were nearly 20 percent shorter than conventionally planned ones (p = 0.002), with the largest time savings in hepatobiliary surgery, where the anatomy is hardest to mentally reconstruct from flat images.