An AG-RetinaNet for Embryonic Blastomeres Detection and Counting

Abstract

Embryo morphology assessment is crucial for determining embryo viability in assisted reproductive technology. Traditional manual evaluation, while currently the primary method, is time-consuming, resource-intensive, and prone to inconsistencies due to the complex analysis of morphological parameters such as cell shape, size, and blastomere count. For rapid and accurate recognition and quantification of blastomeres in embryo images, Attention Gated-RetinaNet (AG-RetinaNet) model is proposed in this article. AG-RetinaNet combines an attention block between the backbone network and the Feature Pyramid Network to overcome the difficulties posed by overlapping blastomeres and morphological changes in embryo shape. The proposed model, trained on a dataset of human embryo images at different cell stages, uses ResNet50 and ResNet101 as backbones for performance comparison. Experimental results demonstrate its competitive performance against state-of-the-art detection models, achieving 95.8% average precision while balancing detection accuracy and computational efficiency. Specifically, the AG-RetinaNet achieves 83.08% precision, 91.13% sensitivity, 90.91% specificity, and an F1-score of 86.92% under optimized Intersection Over Union and confidence thresholds, effectively detecting and counting blastomeres across various grades. The comparison between these results and the manual annotations of embryologists confirms that our model has the potential to improve and streamline the workflow of embryologists in clinical practice.