Morphological Classification of Low Quality Sperm Images Using Deep Learning Networks

The fertility of men and women are examined separately in the diagnosis of infertility. Clinical studies have shown that male infertility rate has a high rate of 25-30% in general diagnosis. Sperm concentration, motility and morphological abnormality are evaluated in male based infertility. In morphological analysis, sperm images should be obtained in detail to obtain objective results. However, the usage of low quality video camera or vibrations occurred in camera module causes to obtain low quality images. In this study, in order to increase the classification performance of the SCIAN-Morpho dataset with low quality sperm images, firstly interpolation methods were applied to increase the data quality. Then, data augmentation techniques have been applied for the data imbalance problem. In the classification phase, pre-trained convolutional neural networks were applied. As a result of the classification, 62% accuracy, 85% precision and 75% sensitivity were obtained by using the VGG-19 networks with the data augmentation and interpolation techniques.