Sperm motility assessed by deep convolutional neural networks into WHO categories

Area of Science:

• Reproductive biology
• Artificial intelligence in medicine
• Medical diagnostics

Background:

• Manual semen analysis for sperm motility is the gold standard but requires extensive training for accuracy.
• Deep convolutional neural networks (DCNNs) excel at image classification tasks, showing potential for automating semen analysis.

Purpose of the Study:

• To evaluate the performance of the DCNN ResNet-50 model in predicting sperm motility proportions according to WHO categories.
• To compare DCNN predictions against manual assessments from reference laboratories.

Main Methods:

• Two DCNN models (3-category and 4-category) were trained using 65 video recordings of semen preparations.
• Tenfold cross-validation was employed, with models trained on mean values from external quality assessment data.
• Performance was assessed using mean absolute error (MAE) and Pearson's correlation coefficient.

Main Results:

• The DCNN models achieved significantly lower error than the baseline.
• Strong correlations were observed between DCNN predictions and manual assessments for progressive motility (r=0.88) and immotile spermatozoa (r=0.89).
• Moderate correlation was found for rapid progressive motility (r=0.673), with variations noted due to interlaboratory differences.

Conclusions:

• DCNN models demonstrate high accuracy in predicting sperm motility categories, comparable to or within interlaboratory variation.
• The AI approach offers a reliable and potentially more efficient method for semen analysis in infertility diagnostics.
• Variability in manual assessments, particularly for rapid progressive motility, highlights challenges in training data standardization.