Effects of Color Space Transformations on Classification Performance of Sperm Morphology


Yüzkat M. , İlhan H. O. , Aydın N.

European Journal of Science and Technology, vol.29, pp.70-75, 2021 (Other Refereed National Journals)

  • Publication Type: Article / Article
  • Volume: 29
  • Publication Date: 2021
  • Journal Name: European Journal of Science and Technology
  • Page Numbers: pp.70-75

Abstract

Infertility is defined by the World Health Organization as the inability of a woman to become pregnant even though the couple had sexual intercourse for one year without any protection. Male and/or female factors might be the reasons for infertility. When diagnosing the male factors, sperm specimens are analyzed in a laboratory environment under certain conditions. The morphological abnormality, characteristic motility and concentration of sperm are examined in the analysis called spermiogram. Spermiogram tests can be done manually by doctors, as well as by using computer-assisted sperm analyzing systems. The importance of computer aided analysis is increasing day by day because visual inspection can give different results from person to person and is costly. In this study, the effect of different color spaces as a preprocessing step is investigated to increase the classification performance of a computer based analyzing approach for sperm morphology. Three sperm morphology data sets abbreviated as SMIDS, HuSHeM and SCIAN-Morpho were used in the experimental tests. Data augmentation was applied on the data sets due to the unbalanced distribution of sperm images among the classes and insufficient data. Then, data sets were converted to two well-known color spaces, LAB and HSV to observe the effects of color space in the classification. MobileNetV2 was utilized as the classification model. In order to indicate the effects of color spaces, results were compared with previously published study where no color transform was implemented. The classification of images in LAB and HSV color spaces had better results than RGB images trained under the same conditions. The maximum classification accuracies of 89%, 85% and 68% were obtained for SMIDS, HuSHeM, SCIAN-Morpho data sets by using the color space transform idea, respectively.