Accurate triage of ovarian masses is facilitated in many centers by the guideline-endorsed, extensively validated IOTA-ADNEX risk model, yet the model still relies on manual measurements of key tumor features. We developed ADNEX-AI, a multi-task deep-learning system that automatically segments four ADNEX ultrasound predictors - lesion, locules, solid tissue, papillary projections - and outputs their quantitative values. The network was trained on 816 annotated images from 369 consecutive women recruited at 11 centers (43% malignancies) and prospectively evaluated on a temporally separate cohort of 1088 patients scanned at 10 of those centers (8008 images; 35% malignancies). ADNEX-AI discriminated benign from malignant tumors with an AUC of 0.930 (95% CI 0.913-0.943), less than but close to examiner-derived ADNEX (0.945; 0.930-0.957; P = 0.004) while delivering better calibration and markedly lower inter-center variability. By removing manual caliper work yet preserving full interpretability, ADNEX-AI could extend high-quality ovarian-cancer risk stratification to clinics that lack specialized ultrasound expertise.
© 2025. The Author(s).