motilitAI

motilitAI predicts human sperm motility by quantifying the percentages of progressive, non-progressive, and immotile spermatozoa from semen sample videos for assessment of sperm quality.

Key Features:

• Dataset: Uses the Visem dataset consisting of videos of human semen samples.
• Input data: Operates on video-based semen samples with per-sperm movement information.
• Tracking: Employs unsupervised tracking to extract movement statistics and displacement features for individual sperm cells.
• Feature representation: Utilizes an aggregated and quantized representation of displacement features for modeling.
• Models: Trains and evaluates multiple neural networks and support vector regression (SVR) models.
• Best model: A linear Support Vector Regressor on aggregated and quantized displacement features achieved the best performance.
• Evaluation metric: Achieved a mean absolute error (MAE) reduction from 8.83 to 7.31 relative to the best Medico Multimedia for Medicine challenge submission on the same dataset and splits.

Scientific Applications:

• Semen quality assessment: Provides automated quantification of progressive, non-progressive, and immotile sperm percentages to assess semen quality.
• Reproductive medicine: Supports clinical and research evaluation of male infertility by supplying quantitative sperm motility metrics.

Methodology:

Processes Visem dataset videos with unsupervised tracking to extract movement statistics and displacement features, then trains neural networks and support vector regression models, with a linear SVR on aggregated and quantized displacement features yielding the best results.