Sperm morphology analysis by using the fusion of two-stage fine-tuned deep networks


Biomedical Signal Processing and Control, vol.71, 2022 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 71
  • Publication Date: 2022
  • Doi Number: 10.1016/j.bspc.2021.103246
  • Journal Name: Biomedical Signal Processing and Control
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, EMBASE, INSPEC
  • Keywords: Sperm morphology analysis, Transfer learning, Decision level fusion, Sperm morphology dataset benchmarking, Deep learning, CLASSIFICATION
  • Yıldız Technical University Affiliated: Yes


© 2021 Elsevier LtdTraditional analysis of sperm morphology refers to the process of extracting and analyzing spatial-based (e.g., sperm eccentricity and sperm area) and/or transform based features (e.g., wavelet decomposition, Fourier analysis and descriptors) from microscopic images, with the ultimate goal being the acquisition of predictive models. Deep learning methods, on the other hand, are capable of learning the features that best distinguish normal and abnormal sperm pixels and can be learned in an end-to-end manner. Application of deep learning in microscopic imaging provides automatized discovery of sperm morphology features as well as exploration of the hierarchy and interaction lying behind the deep features obtained from sperm shapes. In the proposed study, two deep learning networks, which have already been trained by using benchmark image data sets, were employed to classify raw sperm patches by using high level spatial features comprising extreme distinctive ability. Subsequently, a decision level ensemble learning scheme was applied to the predictions of employed deep-nets by using a soft-voting rule. The experiments were carried on three public data sets named as the Sperm Morphology Image Data Set (SMIDS), Human Sperm Head Morphology Set (HuSHeM) and SCIAN-Morpho for performance validation. Additional to classical transfer of pre-trained network weights followed by fine-tuning, which was tested in both data sets, a two-stage sequential fine-tuning procedure was applied to HuSHeM and SCIAN-Morpho resulting in increased accuracy. The results point out that the learning model, in which the predictions of single-stage fine-tuned deep-nets are fused by using soft-voting, performs well for SMIDS, HuSHeM and SCIAN-Morpho reaching 90.87%, 88.89% and 72.08% accuracy values respectively. Besides, the two-stages fine-tuning approach increases the HuSHeM and SCIAN-Morpho performance up to 92.1% and 73.2% without any manual intervention in contrast to competitors employing manual cropping, rotation and biased augmentation steps.