Authors: T. V. Nguyen1 J . M. M. Hall1-3, S. M. Diakiw1, M. D. VerMilyea4,5, A. W. Dinsmore6, D. Perugini1, M. Perugini1,7.
Objective: To investigate whether a non-invasive, deep learning AI algorithm trained on static images of oocytes, denuded prior to ICSI, can predict whether oocytes will develop into a usable blastocyst.
Materials and Methods: This study involved a prospectively collected dataset of 1180 static oocyte images, denuded and imaged immediately prior to ICSI. Images were provided for 116 consecutive patients treated at a single US clinic in 2021. Each image contained a single oocyte, with a linked blastocyst development outcome. Usable blastocyst development of Gardner grade 3BB occurred at a ratio of 41:59, and male infertility factors were also recorded. AI models were developed, both pre- and post-removal of male infertility cases, using oocyte images labeled with usable blastocyst outcome. Training utilized a recently-reported data cleansing method (UDC), with performance evaluated using binary metrics: overall accuracy, sensitivity and specificity.
Predictive power was assessed on a validation set of 236 images prior to application of UDC. Final predictive power with male infertility cases removed was assessed on a validation set of 92 images.
Results: The UDC data cleansing method was applied to the 1180 oocyte images used for AI development. A mean accuracy of 61.8% for constituent AI models was reported on the uncleansed validation set of 236 images, but subsequent removal of known male infertility cases from the validation set increased the accuracy to 63.2%. Interestingly, 31.6% of known male infertility cases were correctly identified and removed by the UDC method, suggesting it is a major source of mislabeling.
To control for the potential impact of male infertility factors on blastocyst development, all images with known male infertility factors were removed prior to applying a second round of data cleansing, removing data that may include additional undiagnosed male infertility cases or other sources of mislabeling (e.g. laboratory error). A mean accuracy of 66.9% was reported for AI models trained on this cleansed dataset. After the second round of UDC, and using two constituent AI models trained on the whole oocyte image, and one model trained only on the segmented zona pellucida region, an ensemble model with high predictive power was obtained. Overall accuracy was 83.7% on the validation set of 92 images, with a sensitivity for predicting usable blastocyst development of 78.8%, and a specificity of 86.4%.
Conclusions: A novel AI algorithm had high predictive power for assessing whether oocytes will develop into a usable blastocyst from single, static oocyte images, denuded prior to ICSI. A UDC data cleansing technique was used to identify and remove potential sources of error, such as male infertility cases.
Impact Statement: There is currently no widely adopted methodology for determining oocyte competency. Such a development could decrease time, cost and unnecessary stress for patients. This study shows it may be possible to develop a robust AI for non-invasive screening of oocytes prior to ICSI or cryopreservation, allowing informed choices regarding oocyte selection or the necessity of further IVF cycles.
References: Coticchio G, Sereni E, Serrao L, Mazzone S, Iadarola I, Borini A. What criteria for the definition of oocyte quality? Ann. N. Y. Acad. Sci. 2004; 1034:132-44. doi: 10.1196/annals.1335.016.
Dakka MA, Nguyen TV, Hall JMM, Diakiw SM, VerMilyea M, et al. Automated detection of poor-quality data: case studies in healthcare. Sci Rep. 2021; 11(1): 18005.
Shi W, Xu B, Wu LM, Jin RT, Luan HB, , et al. Oocytes with a dark zona pellucida demonstrate lower fertilization, implantation and clinical pregnancy rates in IVF/ICSI cycles. PLoS One. 2014; 9(2): e89409. doi: 10.1371/journal.pone.0089409.