The invention discloses an embryo pregnancy result prediction device based on multimodality, and belongs to the field of medical artificial intelligence. Firstly, images of embryos developed to blastocyst stage after in vitro fertilization and corresponding pregnancy results are obtained, three pictures of blastocyst, inner cell mass and trophoblast cells of embryos are obtained, and the pregnancyresults are labeled as labels, and data are labeled as raw data. Then the image is smoothed with Gaussian kernel function to remove part of the noise, and then the image is normalized. The image is then data-augmented as input data. Multimodal method is used to fuse the images of the three images so that the input image contains three evaluation features. the fused image is passed to ResNet-50 for training, that network is optimized according to the target tag, and and is iterated until the training is complete. With the model, three images can be taken before embryo transfer to predict the pregnancy outcome, and the embryo with high success rate can be selected according to the output results, which can improve the final pregnancy success rate.