Deep Learning Based Two-Dimensional Ultrasound for Follicle Monitoring in Infertility Patients

Abstract

Background: A two-dimensional (2D) ultrasound examination is the primary technique for follicle monitoring, but 2D ultrasound follicle monitoring has significant inter- and intra-observer variability in the measurement of follicle diameter. The aim of this study was to propose a novel deep learning-based automated model for accurate 2D ultrasound follicle monitoring and validate the reliability and repeatability in clinical practice. Methods: A prospective trial including 300 infertile women undergoing ovulation induction (single follicle cycles) or controlled ovarian hyperstimulation (multiple follicle cycles) was conducted in the reproductive center. After 2D ultrasound image acquisition, the mean diameter of each targeted follicle was measured using an automated model, experts, and a novice. Designating the expert assessment as the gold standard, the reliability and repeatability of the automated model for single and multiple follicle cycles were evaluated using the intraclass correlation coefficient and Bland-Altman plots. Results: A total of 228 and 1065 follicles from single and multiple follicle cycles, respectively, were included. The accurate recognition rate of follicle boundaries using the automated model was 0.931. The inter-observer variability of follicle mean diameter measurements in single and multiple follicle cycles were 0.970 and 0.984 for the automated model and experts, and 0.965 and 0.978 for a novice and experts, respectively. Bland-Altman plots showed that 95% limits of agreement for follicle diameter measurement between the automated model and experts (−2.02 to 2.39 mm and −0.68 to 1.50 mm) was lower than a novice and experts (−1.69 to 2.74 mm and −0.58 to 1.73 mm) for both single and multiple follicle cycles. The intraclass correlation (ICC) of follicle diameters ≥10 mm calculated by the automated model was significantly higher than follicle diameters <10 mm in multiple follicle cycles (0.834 vs. 0.609). There were no significant differences between the two groups in single follicle cycles (0.967 vs. 0.970). Conclusion: A deep learning-based automated model provides an accurate and fast approach for novices to improve the reliability and receptivity of 2D ultrasound follicle monitoring, especially in multiple follicle cycles and for follicles with a mean diameter ≥ 10 mm.

Keywords

• follicle monitoring
• deep learning
• ultrasound
• image segmentation
• in vitro fertilization