JBRA Assist. Reprod. 2025;29(Suppl 1):27-27
POSTER PRESENTATION
doi: 10.5935/1518-0557.20250081
1Cegyr, Eugin Group
Objective:
• To evaluate the accuracy of AI algorithms in assessing ultrasound images for predicting low ovarian response based on antral follicle count (AFC).
• To compare the performance of AI in predicting low ovarian response via AFC with that of experienced ultrasonographers.
• To assess the capability of AI tools in estimating the number of metaphase II (MII) oocytes on the ovulation triggering day.
• To examine the reliability of AI-based ultrasound evaluations in forecasting ovarian response and MII oocyte count compared to traditional methods.
• To determine the potential role of AI in improving clinical decision-making for ovarian stimulation protocols in ART cycles.
Methods: Prospective cohort study. Conducted between June and December 2024 with 166 participants. Data were collected at the initial ovarian stimulation visit. We included all the patients that underwent an ART cycle (including IVF and oocyte cryopreservation). Egg donors were excluded. Each participant underwent AFC evaluation and trigger evaluation by both an experienced ultrasonographer (EU) and an automated device (Folliscan™, AI). Medical teams received only the traditional ultrasound report. Sensitivity, specificity, and AUC were calculated to predict a low ovarian response (≤4 MII oocytes), with gonadotropin doses ranging from 225-300 IU/day. Statistical analysis included t-test for continuous variables, chi-square for dichotomic variables, McNemar test for sensitivity and specificity comparison and ROC curve analysis to evaluate AUC for both methods.
Results: We analyzed 166 patients (36.7±3.5 years) undergoing ART, with 35.5% (59/166) exhibiting a low ovarian response (≤4 MII oocytes). AFC was lower when measured by EU than AI in both normo- (16.5 vs. 20.0, p<0.01) and poor responders (8.6 vs. 10.0, p<0.01). Both methods predicted poor response, with significantly lower AFC values (EU: 16.4 vs. 8.6, p<0.01; AI: 20.0 vs. 10.0, p<0.01). The optimal cut-off for each technique showed EU sensitivity (Se) 88.1% and specificity (Sp) 69.2% (area under the curve -AUC- 0.87; 95% CI 0.81-0.93), while AI had Se 83.1% and Sp 68.2% (AUC 0.80; 95% CI 0.73-0.87), with no significant difference between techniques (p=0.71). For ovulation trigger monitoring, both methods identified more large follicles in normothan in poor responders (EU: 5.2 vs. 2.8, p<0.01; AI: 4.1 vs. 1.6, p<0.01), with EU detecting 1.1 more follicles than AI (p<0.01). EU demonstrated slightly greater accuracy in predicting retrieved MII oocytes (59% vs. 66%, p<0.01).
Conclusion: AI-based ultrasound could be a valuable alternative in settings lacking experienced ultrasonographers. While not assessed for all monitoring stages, its AFC performance matched expert assessments, aiding ovarian stimulation planning. However, for ovulation triggering, it may underestimate MII oocytes, requiring further data for validation and improved accuracy. The ovulation trigger decision was based on traditional follicle count, potentially influencing results if the automated measurement had been used instead. This pilot study requires validation with a larger sample size.