MADRNet: Morphology-Aware Dual-Path Reversible Network for Sperm Classification.

Area of Science:

• Reproductive Biology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual sperm morphology analysis is subjective and impacts male infertility diagnosis accuracy.
• Existing automated sperm classification models often require manual intervention and lack alignment with WHO standards.
• Clinical assessment relies on biomarkers like head aspect ratio and acrosomal integrity.

Purpose of the Study:

• To develop an automated, accurate, and efficient sperm morphology analysis model aligned with WHO standards.
• To overcome the subjectivity and inconsistencies of manual sperm evaluation.
• To integrate key clinical biomarkers into an AI framework for improved male infertility diagnosis.

Main Methods:

• Proposed a morphology-aware dual-path reversible network (MADRNet) incorporating head aspect ratio and acrosomal integrity.
• Utilized a dual-path attention mechanism (spatial and channel attention) with acrosome anatomical constraint.
• Developed a dynamic loss function for head aspect ratio constraint and employed a reversible architecture.
• Trained and evaluated the model on the HuSHeM dataset.

Main Results:

• MADRNet achieved 96.3% accuracy and 96.8% F1 score on the HuSHeM dataset.
• The model demonstrated real-time processing at 32ms per image.
• The reversible architecture preserved microscopic details while reducing GPU memory usage.

Conclusions:

• MADRNet offers a precise and efficient automated solution for clinical sperm screening.
• The model effectively addresses limitations of manual analysis and existing automated methods.
• The developed AI tool shows significant potential for improving male infertility diagnosis.