Computer vision-based artificial intelligence systems for real-time identification of significant intraoperative events

Computer vision-based artificial intelligence (CV-AI) for intraoperative use has advanced rapidly, yet translation into routine surgical practice remains limited. This thesis examined what is required for surgical CV-AI to move from conceptual promise to defensible real-world intraoperative implementation.A multi-stage implementation-focused programme of research was undertaken in laparoscopic cholecystectomy. First, a scoping review mapped the contemporary general surgical CV-AI literature, performance metrics, and evidence of readiness for real-time deployment. Second, a surgeon-blinded in-theatre design and integration pilot evaluated feasibility, workflow safety, operational stability, and early usability of a context-aware multi-task CV-AI platform. Third, prospective live deployment in a real-world cohort quantified real-time actionability of outputs related to significant intraoperative events, evaluated downstream task performance, and characterised dominant failure mechanisms using integrated mixed-methods analysis. Finally, surgeon attitudes before and after deployment were assessed to identify the adoption conditions under which intraoperative CV-AI may become acceptable in practice, including perceived usefulness, autonomy, accountability, and governance.The scoping review confirmed that the contemporary literature remains dominated by retrospective feasibility studies, with limited real-time evaluation and substantial heterogeneity in reporting. In-theatre deployment was feasible and workflow safe, achieving high recording completeness without any observed device-related safety events, and remediable system failures were identified. Prospective live evaluation showed that real-time CV-AI could generate clinically relevant intraoperative signals, but practical utility was constrained more by timeliness and predictable context-dependent failure modes than by headline accuracy alone. Performance degradation clustered around limited visual access, workflow variation, proxy triggering, and construct mismatch. Surgeon attitudes reflected conditional receptivity, with sustained willingness to use CV-AI alongside persistent concerns regarding liability, governance, and preservation of end-user autonomy.This thesis demonstrated that implementation readiness for surgical CV-AI is determined not by algorithm discrimination performance alone, but by task-specific utility, workflow robustness, interpretability, and governance. These findings support a staged translational pathway from surgeon-blinded feasibility research to prospective surgeon-visible evaluation, with early prioritisation of low-risk, documentation-oriented applications before progression to higher-stakes decision support.