Effective use of prior information for high-performance embryo grading

Abstract

The selection of top-quality embryos is the key step for a successful outcome of in-vitro fertilization. The existing CNN-based methods directly use a classical network to grade/classify embryos and discard the potentially useful prior information of cells. Thus, we used the features of cell boundaries as the effective representation of cells and proposed a two-stage network to integrate features of cell boundaries with those of embryo for improving embryo grading performance. The first-stage network was a segmentation network. It extracted the features of cell boundaries as the effective and implicit representation of cells, which mainly determined the grades of embryos in clinical practice. The combination of different-level boundary features was input into the second-stage network to boost its discriminative ability for embryo grading. The second-stage network was a classification network. It integrated the features of cell boundaries with those of embryo. The integration of features enables the second-stage network to learn more useful information for embryo grading. Compared with the second-stage classification network alone, our two-stage network with one set of boundary features achieved higher embryo grading performance, and our network with two sets of boundary features further improved the performance. The two-stage network made full use of cell boundaries as the effective and implicit representation of prior information, and integrated the features of cell boundaries with the extracted classification features to improve the embryo grading performance.