Dual Stream Deep Neural Network for Joint Sperm Morphology and Motility Estimation

Morphology and motility are vital features to evaluate sperm quality, which is essential for estimating the success of assisted reproduction. Current standards from the World Health Organization (WHO) suggest that human operators are needed to perform sperm quality assessments. However, human assessment is proven to be subjective and leads to variability in the results. To cope with that problem, an automated approach is needed. This study aimed to perform a joint motility and morphology estimation via a single model that has never been done before. The model takes a stack of optical flow images and a grayscale image to be processed by motion and appearance stream. The outputs from each stream are combined and further processed by a regressor to estimate the morphology and motility. The proposed approach performed better than similar methods by achieving mean average errors of 4. 4 9 5 % and 7. 8 6 5 % for morphology and motility estimation, respectively.