Design: Retrospective cohort study.

Setting: French university hospital (single-center).

Patients or Participants: 181 patients who underwent laparoscopy for suspected endometriosis and had a pelvic MRI less than 15 months before surgery between 2019 and 2021.

Interventions: The cohort was divided into a training set (N=135) and a validation set (N=46) for the creation of the artificial intelligence model. Simultaneously, 5 surgeons from the department retrospectively and blindly attempted to predict the laparoscopy outcome.

Measurements and Main Results: In the validation set, the AI model had the following diagnostic performance: Sensitivity (Se) 0.68, Specificity (Sp) 0.81, Positive Predictive Value (PPV) 0.64, Negative Predictive Value (NPV) 0.67, Accuracy 0.65, and Area Under the Curve (AUC) of the ROC curve 0.745. In contrast, the department's surgeons had the following combined performance: Se 0.59, Sp 0.65, PPV 0.80, NPV 0.35, Accuracy 0.58, AUC 0.61. The AUC of the ROC curve for the AI model was significantly higher than that of human prediction (p=.031). Factors significantly associated with the presence of retrocervical endometriosis were: age, history of endometriosis, history of endometriosis surgery, dysmenorrhea, presence of deep endometriosis lesion on ultrasound, presence of left USL or digestive lesion on MRI.

Conclusion: The AI model appears to have higher diagnostic performances compared to human analysis. This model requires internal and external validation but could lead to significant improvement in the management of patients with deep retrocervical endometriosis.