Interpretable Sperm Morphology Classification via Attention-Guided Deep Learning

This study introduces an attention-guided deep learning framework for sperm morphology classification, integrating a pretrained EfficientNet-B0 with a Convolutional Block Attention Module (CBAM) to enhance interpretability. The model achieves accuracy rates of 90.2% and 93.9% on the SMIDS and HuSHem datasets, respectively, while outperforming SimpleCNN and standard EfficientNet-B0, and employs Grad-CAM++ for visualizing influential features. This framework addresses the critical need for interpretable AI in clinical fertility assessments, potentially increasing the adoption of automated sperm analysis tools.