Evaluation of machine learning models for classification of fertilization status of African clawed frog (xenopus laevis) by use of synthetic images

The African clawed frog ( Xenopus laevis ) has been a pivotal model organism providing valuable insights into cell biology, developmental biology, and human disease. Fast and accurate identification of fertilized embryos versus unfertilized X. laevis eggs is crucial for time sensitive experimental applications such as assessing the efficacy of sperm cryopreservation and artificial fertilization. Traditionally, this task has relied on manual annotation, a labor-intensive process. To address this challenge, we developed a convolutional neural network (CNN)-based tool using the YOLOv10b (“You only look once version 10 balanced accuracy”) architecture to detect and quantify fertilized and unfertilized eggs in images. Our tool leveraged synthetic data generation to automate the creation of 1,194 training and validation images and annotations, significantly reducing the need for manual annotation. Through testing on 120 synthetic and 20 experimental images, we evaluated the performance of three optimizer models: Adam, AdamW, and SGD. The accuracy when testing the synthetic data was 87% for the Adam model, 89% for the AdamW model, and 90% for the SGD. On the other hand, the median percentage embryo classification accuracy (ECA) values were 97% (Adam), 95% (AdamW), and 96% (SGD). Our results demonstrate the feasibility of rapid development of machine learning tools for classification of fertilization making it a valuable tool for improving the efficiency in laboratory operation for biomedical researchers and resource centers that use X. laevis as a model organism. The codes to image synthesizing, training, and testing along with the Graphical User Interface (GUI) were made publicly available on GitHub and Open Science Framework (OSF).