An artificial intelligence-based algorithm for predicting pregnancy success using static images captured by optical light microscopy during intracytoplasmic sperm injection

• Päätekijät: Jared Geller (Tekijä), Ineabelle Collazo (Tekijä), Raghav Pai (Tekijä), Nicholas Hendon (Tekijä), Soum D Lokeshwar (Tekijä), Himanshu Arora (Tekijä), Manuel Molina (Tekijä), Ranjith Ramasamy (Tekijä)
• Aineistotyyppi: Kirja
• Julkaistu: Wolters Kluwer Medknow Publications, 2021-01-01T00:00:00Z.
• Aiheet: article
• Linkit: Connect to this object online.
• Notes: 0974-1208; 1998-4766; 10.4103/jhrs.jhrs_53_21
• Keywords: deep learning; in vitro fertilization; inception v1; transfer learning; Gynecology and obstetrics; RG1-991
• Source: Journal of Human Reproductive Sciences, Vol 14, Iss 3, Pp 288-292 (2021)

Context (Background): Analysis of embryos for in vitro fertilization (IVF) involves manual grading of human embryos through light microscopy. Recent research shows that artificial intelligence techniques applied to time lapse embryo images can successfully ascertain embryo quality. However, laboratories often capture static images and cannot apply this research in a real-world setting. Further, current models do not predict the outcome of pregnancy. Aims: To create and assess a convolutional neural network to predict embryo quality using static images from a limited dataset. We considered two classification problems: predicting whether an embryo will lead to a pregnancy or not and predicting the outcome of that pregnancy. Settings and Design: We utilized transfer learning techniques using a pretrained Inception V1 network. All models were built using the Tensorflow software package. Methods: We utilized a total of 361 randomly sampled static images collected from four South Florida IVF clinics. Data were collected between 2016 and 2019. Statistical Analysis Used: We utilized deep-learning techniques, including data augmentation to reduce model variance and transfer learning to bolster our limited dataset. We used a standard train/validation/test dataset split to avoid model overfitting. Results: Our algorithm achieved 0.657 area under the curve for predicting pregnancy versus nonpregnancy. However, our model was unable to meaningfully predict whether a pregnancy led a to live birth. Conclusions: Despite the limited dataset that achieved somewhat of a lower accuracy than conventional embryo selection, this is the first study that has successfully made IVF predictions from static images alone. Future availability of more data may allow for prospective validation and further generalisability of results.