Impact of transfer learning for human sperm segmentation using deep learning.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Reproductive Medicine

Background:

• Male infertility affects nearly half of infertile couples, necessitating accurate semen analysis.
• Traditional sperm morphology evaluation is prone to errors, highlighting the need for automated methods.
• Improving sperm segmentation precision is crucial for reliable computer-assisted sperm analysis.

Purpose of the Study:

• To assess the utility of two deep learning models (U-Net and Mask R-CNN) for segmenting human sperm heads, acrosomes, and nuclei.
• To enhance the precision of automated sperm morphology analysis for clinical applications.

Main Methods:

• Utilized U-Net and Mask R-CNN deep learning architectures for sperm part segmentation.
• Employed data augmentation, cross-validation, hyperparameter tuning, and transfer learning.
• Evaluated models on the SCIAN-SpermSegGS public dataset.

Main Results:

• U-Net with transfer learning achieved high Dice coefficients for sperm head (0.96), acrosome (0.94), and nucleus (0.95) segmentation.
• These results surpassed existing state-of-the-art sperm part segmentation methods.
• Transfer learning substantially improved segmentation accuracy and reduced failure cases.

Conclusions:

• Deep learning, particularly U-Net with transfer learning, offers a promising approach for accurate sperm morphology analysis.
• This advancement can lead to more reliable computer-assisted diagnosis of male infertility.
• Further development in sperm segmentation is key to clinical adoption.