Refined trajectory smoothing and deep learning classification of human sperm motility

Area of Science:

• Reproductive biology andrology
• Artificial intelligence in medicine
• Biomedical signal processing

Background:

• Conventional computer-aided sperm analysis (CASA) systems often use basic smoothing algorithms for motility parameter extraction, leading to inaccuracies in key metrics like beat cross frequency (BCF) and amplitude of lateral head displacement (ALH).
• These inaccuracies can result in misclassification of sperm, particularly for identifying hyperactivated spermatozoa, which is crucial for assisted reproductive technologies (ART).

Purpose of the Study:

• To investigate if precise trajectory smoothing can enhance the extraction of sperm motility features.
• To determine if deep learning models trained on raw sperm trajectory data can accurately classify different sperm motility patterns, such as progressive and hyperactivated.

Main Methods:

• Compared Gaussian Process Regression (GPR), moving average, and Discrete Cosine Transform (DCT) smoothing for average path estimation using 2326 sperm trajectories.
• Introduced a novel metric, path average width (PAW), to quantify lateral head displacement.
• Developed an ensemble of InceptionTime deep learning models trained on raw (x, y) coordinate sequences for sperm classification.

Main Results:

• Discrete Cosine Transform (DCT) smoothing with 12 frequency components (DCT-12) yielded the most consistent average paths, improving BCF and ALH accuracy.
• The novel PAW metric effectively differentiated hyperactivated (5.5 ± 1.5 μm) from progressive spermatozoa (2.0 ± 1.3 μm).
• The InceptionTime model achieved 89% accuracy in classifying progressive versus hyperactivated trajectories and 78% accuracy for motility grades.

Conclusions:

• The proposed approach enhances sperm motility parameter extraction precision using frequency-domain smoothing and enables accurate classification of sperm motility patterns via deep learning on raw trajectory data.
• This integrated method addresses core limitations in CASA systems and shows potential for real-time application in ART workflows.
• Future work should focus on training models with high-viscosity media data and multi-centre validation for improved clinical translation and generalizability.