Artificial Intelligence for Surgical Scene Understanding: A Systematic Review and Reporting Quality Meta-Analysis

Surgical scene understanding (SSU) describes the use of Artificial Intelligence (AI) to provide an understanding of visual components of surgical imaging data, such as laparoscopic surgery videos. While hundreds of publications report AI capabilities to identify instruments, anatomical structures, and other contextual data and testify potential for real-time support in the operating room, the clinical implementation of SSU remains limited.

This systematic review and meta-analysis (registered in the PROSPERO database under CRD420251005301) assesses the current state and research gaps in computational SSU, focusing on data curation, model design, validation, uncertainty estimation, performance metrics, reporting quality, and clinical applicability. Studies were included if they analyzed intraoperative data from minimally invasive abdominal surgeries in humans, developed computational SSU methods, and reported trainable models with formal validation and performance metrics.

A total of 188 studies from six literature databases were included. Most relied on small, single-center datasets, often from laparoscopic cholecystectomies, with limited metadata and topical diversity. Research was largely descriptive, with limited reporting on clinical relevance, limitations, code availability, and model uncertainty. Validation was often inadequate, typically relying on simple hold-out strategies, with limited testing on external datasets and purely technical validation approaches without any clinical expert involvement. Clinical translation was addressed in only eleven works. Overall, studies showed minimal progress toward real-world application. Our findings highlight the need for diverse, multi-institutional datasets, robust validation practices, and clinically driven development to unlock the full potential of SSU in surgical practice.