A Deep Learning Framework Design for Automatic Blastocyst Evaluation With Multifocal Images

To design a deep learning network model for automatic human blastocyst quality assessment with multifocal images, a total of 11,275 images of 1,025 normally fertilized blastocysts underwent traditional in vitro fertilization (IVF) treatment between May 2017 and August 2018. Images captured 116 +/- 1 hours after insemination were classified according to embryo grade as follows: good, poor. Multiple images with different focal depths were designed as a batch input to the proposed deep learning network so that the network had more information about the blastocysts and higher accuracies were expected. In our dataset, the area under the curve (AUC) of the best model reached 0.936. Using the Grad-CAM algorithm to visualize all blastocyst stage images on the test set, it was found that key features relied on by the classification model were trophectoderm (TE), inner cell mass (ICM) and zona pellucida (ZP). A deep learning approach provides a new method for assessing blastocyst quality. The multichannel combination network showed the best performance among the three proposed models, in which the clearest images from multiple focal depths were chosen and fetched into the VGG-16 network as different channels.