Deep learning-based sperm motility and morphology estimation on stacked color-coded MotionFlow

Abstract

Motility and morphology are crucial factors in determining male fertility. The current gold standard defined by the World Health Organization (WHO) mandates that semen analysis be performed by trained technicians. Despite strict standardization and technical guidelines set by the WHO, variability in semen analysis results remains prevalent owing to human subjectivity. Computer-Aided Sperm Analysis presents a further challenge because of its poor agreement with human analysis. This study aimed to enhance the accuracy of automated semen analysis by introducing a new method for expressing sperm motion and investigating advanced deep neural network architectures to estimate motility and morphology. Initially, we extracted motion information from the VISEM dataset using our novel motion representation technique called MotionFlow, along with shape information. Consequently, motility and morphology neural networks were constructed to exploit transfer learning in other fields to improve performance. These networks ingested motion and shape features and made separate predictions for motility and morphology. The evaluation process utilized a K-Fold cross-validation scheme to determine the mean absolute error (MAE) and maintain objectivity throughout the analysis. The proposed method achieved a greater level of performance than the current methods by attaining MAE of 6.842% and 4.148% for motility and morphology estimation, respectively. The improvement accomplished by this research may pave the way toward a fully automated human sperm quality assessment.