Artificial-intelligence assistance for diagnosis, detection and surgical decision-making in colorectal peritoneal metastases: A cross-sectional BIG-RENAPE pilot study.

BACKGROUND

Human visual performance in diagnosing and detecting peritoneal metastases (PM) during laparoscopy remains limited. A deep-learning algorithm was developed to assist PM detection on laparoscopic videos. This study evaluated its impact on diagnostic accuracy, detection capability, and therapeutic decision-making.

METHODS

Laparoscopic video sequences showing 14 nodules and 20 selected peritoneal regions were randomly extracted from a large dataset. An online survey was distributed to digestive surgeons specialized in oncologic surgery, either experts in PM management (BIGRENAPE) or non-experts (non-BIGRENAPE). For each nodule, surgeons assessed malignancy (benign vs malignant) and selected a management strategy: biopsy vs no biopsy for non-BIGRENAPE, and resection/destruction, frozen-section biopsy, or no action for BIGRENAPE, first without then with AI assistance. For each region, surgeons also estimated the number of malignant nodules with and without AI assistance.

RESULTS

Sixty-nine surgeons participated (44 non-BIGRENAPE, 25 BIGRENAPE). Diagnostic accuracy with AI assistance remained stable, from 63% to 61% in non-BIGRENAPE and from 70% to 71% in BIGRENAPE experts. Among non-BIGRENAPE surgeons, correct decision-making remained stable from 67% to 70%. In contrast, AI significantly enhanced appropriate surgical decision-making among BIGRENAPE experts from 65% to 69% (p = 0.03; OR 2.27, 95% CI 1.08-5.11) and modified therapeutic strategy in 20.9%. Detection performance showed a moderate correlation between true and estimated malignant nodule counts (Spearman 0.72-0.83), with slight deterioration under AI assistance.

CONCLUSIONS

Human performance in diagnosing, detecting, and managing colorectal PM remains suboptimal. AI assistance improved expert surgical decision-making but not diagnostic or quantification performance.