Research question

Can a deep machine learning artificial intelligence algorithm predict ploidy and implantation in a known data set of static blastocyst images, and how does its performance compare against chance and experienced embryologists?

Design

A database of blastocyst images with known outcome was applied with an algorithm dubbed ERICA (Embryo Ranking Intelligent Classification Algorithm). It was evaluated against its ability to predict euploidy, compare ploidy prediction against randomly assigned prognosis labels and against senior embryologists, and if it could rank an euploid embryo highly.

Results

A total of 1231 embryo images were classed as good prognosis if euploid and implanted or poor prognosis if aneuploid and failed to implant. An accuracy of 0.70 was obtained with ERICA, with positive predictive value of 0.79 for predicting euploidy. ERICA had greater normalized discontinued cumulative gain (ranking metric) than random selection (P = 0.0007), and both embryologists (P = 0.0014 and 0.0242, respectively). ERICA ranked an euploid blastocyst first in 78.9% and at least one euploid embryo within the top two blastocysts in 94.7% of cases, better than random classification and the two senior embryologists. Average embryo ranking time for four blastocysts was under 25 s.

Conclusion

Artificial intelligence lends itself well to image pattern recognition. We have trained ERICA to rank embryos based on ploidy and implantation potential using single static embryo image. This tool represents a potentially significant advantage to assist embryologists to choose the best embryo, saving time spent annotating and does not require time lapse or invasive biopsy. Future work should be directed to evaluate reproducibility in different data sets.

中文翻译：

---

Embryo Ranking Intelligent Classification Algorithm (ERICA)：预测胚胎倍性和着床的人工智能临床助手。

研究问题

深度机器学习人工智能算法能否在已知的静态囊胚图像数据集中预测倍性和植入，以及它的性能与机会和经验丰富的胚胎学家相比如何？

设计

将具有已知结果的囊胚图像数据库与称为 ERICA（胚胎排名智能分类算法）的算法一起应用。它根据其预测整倍体的能力进行评估，将倍性预测与随机分配的预后标签和高级胚胎学家进行比较，以及是否可以对整倍体胚胎进行高排名。

结果

共有1231个胚胎图像被归类为整倍体和植入的预后良好或非整倍体和未植入的预后不良。ERICA 的准确率为 0.70，预测整倍体的阳性预测值为 0.79。ERICA 比随机选择（P  = 0.0007）和胚胎学家（ 分别为P = 0.0014 和 0.0242）具有更大的标准化中断累积增益（排名指标）。ERICA 以 78.9% 的比例将整倍体囊胚排在第一位，在 94.7% 的案例中，前两个囊胚中至少有一个整倍体胚胎，优于随机分类和两位资深胚胎学家。四个囊胚的平均胚胎排序时间低于 25 秒。

结论

人工智能非常适合图像模式识别。我们已经训练 ERICA 使用单个静态胚胎图像根据倍性和着床潜力对胚胎进行排名。该工具具有潜在的显着优势，可帮助胚胎学家选择最佳胚胎，节省注释时间，并且不需要延时或侵入性活检。未来的工作应针对评估不同数据集的再现性。