Adaptive Logit Fusion for Mitigating Class Imbalance in Multi-Category Sperm Morphology Assessment

Area of Science:

• Reproductive Biology
• Computer Science
• Medical Imaging

Background:

• Sperm morphology is a key factor in male fertility assessment.
• Accurate classification of sperm morphology is crucial for diagnosis.
• Existing methods may struggle with the complexity and class imbalance of sperm morphology data.

Purpose of the Study:

• To develop and evaluate a deep learning-based automated system for classifying sperm morphology into 18 classes.
• To improve classification accuracy and generalization, especially under class imbalance.
• To investigate the effectiveness of an ensemble strategy combining state-of-the-art convolutional neural networks.

Main Methods:

• Utilized EfficientNetV2-S and ResNet50V2 convolutional neural networks (CNNs).
• Employed fine-tuning with class-weighted loss functions and extensive data augmentation.
• Implemented automatic mixed precision training and an ensemble strategy by fusing model logits.
• Optimized fusion weights to maximize recall, precision, and F1-score.

Main Results:

• The proposed ensemble model achieved an overall accuracy of 70.94%, outperforming individual CNN models.
• High accuracy was observed for classifying sperm with distinct abnormalities like PinHead and DoubleTail.
• Performance was comparatively lower for sperm with less visually distinctive defects.

Conclusions:

• Deep learning-based ensemble models can provide consistent and reliable automated sperm morphology classification.
• The developed approach shows potential for enhancing male fertility diagnostics.
• Further refinement may be needed for classifying subtle morphological abnormalities.