EM-aided Federated Learning for Quality Assessment of Embryo Development

The issue of leveraging federated learning to improve data privacy in IVF embryo selection is addressed. Multiple deep learning models are incorporated with the EM algorithm to form a federated learning framework for the quality evaluation of embryo cleavage using 2D images. The framework comprises plural clients and a server, characterized in that each client is EM-trained with local data. Upon the completion of the local training, a separate 5-mode distribution is generated for each client. The clients’ distribution statistics are then uploaded to the server and aggregated to produce a global distribution. During inference, each client uses image classifiers and an instance segmentor, with the global distribution acting as a reference calibrator to (1) identify the absolute cleavage timing, i.e., tPNa, tPNf, t2, t3, t4, t5, t6, t7, and t8, (2) track the cleavage process of embryo to detect the irregular cleavage, and (3) assess the symmetry degree of blastomeres. Such an EM-aided learning facilitates the adaptability of classifiers and the segmentor to bias arising from data collection at different locations or countries. Experimental results show that the proposed framework outperforms commercial Time-Lapse Incubators in reducing the average error of timing prediction.