3D Convolutional Neural Networks for Sperm Motility Prediction

Semen analysis is an important analysis for male infertility primary investigation. Sperm motility is one of the main indicators for pregnancy and conception rate, and it could be classified into three motility groups which are progressive, non-progressive and immotile spermatozoa according to WHO manual. Manual semen analysis has been revealed with accuracy and precision limitation due to noncompliance to guidelines and procedures. On the other hand, the commercialized automated semen analyzer is not recommended for clinical use due to their analysis results not comparable with manual methods. Their handling procedures received criticisms as the proper guidelines were not discussed and reviewed by WHO. In this study, we aim to employ deep learning methods for sperm motility prediction using three-dimensional CNN (3DCNN). Firstly, datasets are prepared by extracting dense optical flow frames with different stride number from semen videos and stacked together forming 3D input. Next, a 3DCNN was designed to adopt stacked dense optical flow frames and the results obtained using datasets generated with different stride number were compared and analysed. As a result, 3DCNN has better accuracy compared with other deep learning approaches explored by other similar research works with average mean absolute error of 8.506. The source code for this research work is made public at Github repository: https://github.com/GohVh/3DCNN-SpermMotilityPrediction.