MFIF-Net: A Multi-Focal Image Fusion Network for Implantation Outcome Prediction of Blastocyst

Abstract

Accurately predicting implantation outcomes based on blastocyst developmental potential is valuable in in-vitro fertilization (IVF). Clinically, embryologists analyze multiple focal-plane images (FP-images) to comprehensively assess embryo grades, which is extremely cumbersome and easily prone to inconsistency. Developing automatic computer-aided methods for analyzing embryo images is highly desirable. However, effectively fusing multiple FP-images for prediction remains a largely under-explored issue. To this end, we propose a novel Multiple Focal-plane Image Fusion Network, called MFIF-Net, to predict implantation outcomes of blastocyst. Specifically, our MFIF-Net consists of two sub-networks: a Core Image Generation Network (CI-Gen) and a Key Feature Fusion Network (KFFNet). In CI-Gen, we fuse multiple FP-images to generate a core image by pixel-wise weighting since different FP-images can have different focus positions. To further capture key features in each FP-image, we propose KFFNet to extract key information from the FP-images again and fuse them with the core image. In KFFNet, a Fusion Module is designed to capture key information of each FP-image, for which Squeeze Multi-Headed Attention is developed to exchange features and mitigate computationally intensive issues in attention. Comprehensive experiments validate the superiority and the rationality of our MFIF-Net approach over state-of-the-art methods in various metrics. Ablation studies also confirm the positive impact of each component in our MFIF-Net.