EP13.17: Prospective validation of imaging‐based endometriosis staging systems: AAGL‐US, UBESS and ultrasound based rASRM

Résumé fourni par la source

Results: We had included 54 videos from 51 patients with 188 labeled frames.Of these, 171/188 frames contained an annotated DE lesion, and 17/188 were without lesion.On lesion-present frames, Dice was 0.45 ± 0.28 (mean ± SD) with a median of 0.53 (IQR 0.23-0.68;max 0.92).Frame-level overlap was observed in 82% (140/171) of lesion-present frames.Patient-level sensitivity was 90% (46/51) using the ''any-overlap'' definition (5/51 patients had no overlap on any annotated lesion frame).Conclusions: An AI model trained from expert annotations in offline TVS cine clips showed encouraging sensitivity and adequate segmentation agreement for DE affecting the torus uterinus.Larger studies with more control cases are recommended before clinical deployment as a real-time alerting.