Embryo Evaluation Based on ResNet with AdaptiveGA-optimized Hyperparameters

Embryo Evaluation Based on ResNet with AdaptiveGA-optimized Hyperparameters

In vitro fertilization (IVF) embryo evaluation based on morphology is an effective method to improve the success rate of transplantation. Although convolutional neural networks (CNNs) have made great achievements in many image classifications, there are still great challenges in accurately classifying embryos due to the insufficient samples, interference of exfoliated cells, and inappropriate hyperparameter configuration in the classification network. In this paper, a residual neural network optimized by the adaptive genetic algorithm is proposed to evaluate embryos. Firstly, a novel algorithm for extracting the region of interest (ROI) is embedded in the preprocessing part of the model to eliminate exfoliated cells close to the embryo. Secondly, several kinds of specific transformation methods are established to expand the dataset based on the symmetry of embryos. In addition, an adaptive genetic algorithm is adopted to search for optimal hyperparameters. Experiments on the data set provided by Shanghai General Hospital show that the algorithm has an excellent performance in embryo evaluation. The accuracy of our model is 86.4%, the recall is 88.4%, and the AUC is 0.93. Our results indicated that the proposed model can effectively improve the classification performance of ResNet, and thus achieve the clinic requirements of embryo evaluation.