Multi-model CNN fusion for sperm morphology analysis


YÜZKAT M. , İLHAN H. O. , AYDIN N.

COMPUTERS IN BIOLOGY AND MEDICINE, vol.137, 2021 (Journal Indexed in SCI) identifier identifier identifier

  • Publication Type: Article / Article
  • Volume: 137
  • Publication Date: 2021
  • Doi Number: 10.1016/j.compbiomed.2021.104790
  • Title of Journal : COMPUTERS IN BIOLOGY AND MEDICINE
  • Keywords: Sperm morphology, Convolutional neural network (CNN), Data augmentation, Decision level fusion, CONVOLUTIONAL NEURAL-NETWORKS, GOLD-STANDARD, SEMEN, CLASSIFICATION, MOTILITY, QUALITY, HEAD, SEGMENTATION, ACROSOME

Abstract

Infertility is a common disorder affecting 20% of couples worldwide. Furthermore, 40% of all cases are related to male infertility. The first step in the determination of male infertility is semen analysis. The morphology, concentration, and motility of sperm are important characteristics evaluated by experts during semen analysis. Most laboratories perform the tests manually. However, manual semen analysis requires much time and is subject to observer variability during the evaluation. Therefore, computer-assisted systems are required. Additionally, to obtain more objective results, a large amount of data is necessary. Deep learning networks, which have become popular in recent years, are used for processing and analysing such quantities of data. Convolutional neural networks (CNNs) are a class of deep learning algorithm that are used extensively for processing and analysing images. In this study, six different CNN models were created for completely automating the morphological classification of sperm images. Additionally, two decision-level fusion techniques namely hard-voting and softvoting were applied over these CNNs. To evaluate the performance of the proposed approach, three publicly available sperm morphology data sets were used in the experimental tests. For an objective analysis, a crossvalidation technique was applied by dividing the data sets into five sub-sets. In addition, various data augmentation scales and mini-batch analysis were employed to obtain the highest classification accuracies. Finally, in the classification, accuracies 90.73%, 85.18% and 71.91% were obtained for the SMIDS, HuSHeM and SCIAN-Morpho data sets, respectively, using the soft-voting based fusion approach over the six created CNN models. The results suggested that the proposed approach could automatically classify as well as achieve high success in three different data sets.