Deep feature engineering for accurate sperm morphology classification using CBAM-enhanced ResNet50.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Biology
• Reproductive Medicine

Background:

• Sperm morphology analysis is crucial for male fertility assessment.
• Manual analysis is time-consuming, subjective, and has high inter-observer variability.
• Abnormal sperm morphology correlates with reduced fertility and poor ART outcomes.

Purpose of the Study:

• To develop an automated, objective sperm morphology classification framework.
• To combine deep learning (ResNet50, CBAM) with deep feature engineering (DFE).
• To overcome limitations of traditional manual sperm analysis.

Main Methods:

• A hybrid deep learning architecture integrating ResNet50 and CBAM.
• Comprehensive deep feature engineering pipeline with 10 feature selection methods.
• Classification using Support Vector Machines and k-Nearest Neighbors on benchmark datasets (SMIDS, HuSHeM).

Main Results:

• Achieved high test accuracies: 96.08% (SMIDS) and 96.77% (HuSHeM).
• Significant improvements over baseline CNN performance (8.08% and 10.41% increases).
• Outperformed state-of-the-art methods, including Vision Transformers and ensemble methods.

Conclusions:

• Attention-based deep learning with DFE enhances sperm morphology analysis.
• Framework offers standardized, objective fertility assessment, reducing variability.
• Enables significant time savings for embryologists and improves reproducibility, enhancing patient care in reproductive medicine.