Automated sperm morpshology testing using artificial intelligence

Abstract.Abstract1: The standard methods for analyzing sperm morphology using Krueger strict grading criteria is time-consuming and is the main reason most semen analyses take >1 day to finalize. All proposed alternative technologies have either been too expensive or inaccurate for clinical cost-effectiveness. Thus, manual image-based sperm morphology assessment continues to be the gold standard modality in clinical semen analysis. A reliable, inexpensive, automated technology for sperm morphology testing can significantly improve clinical workflow for semen analysis. We developed an artificial neural network, trained and validated with over 3,500 images of sperm acquired using clinical benchtop microscopes. To evaluate the developed artificial intelligence, we allowed the network to evaluate stained slide images that were available to us through the American Association of Bioanalysts (AAB) Proficiency Testing Service (PTS). The developed algorithm assessed a total of 9 semen samples on our system, evaluating at least 350 sperm images per sample. We transfer learned a deep-convolutional network after replacing the final classification layer and retraining it with our dataset of sperm images (1). Images of 3820 sperm were split into training and validation sets with a split of 80-20. The sperm images used in the training were all annotated by 10 different lab technicians with each image annotated thrice. The mode of each image’s annotation was used during the training process to reduce the subjectivity of annotations. We evaluated the network first in its ability in evaluating individual sperm, then we tested our network using 9 slides from the AAB PTS database and compared our network’s output to the global average. When our network was tested with 415 individual sperm images, our network was able to correctly identify 371 sperm (∼89%) based on annotations obtained from technicians. We then evaluated our network’s performance in assessing semen samples for sperm morphology testing. The network performed with an accuracy of 100% in identifying all abnormal and normal samples (n=9) in comparison to the national average reported by the AAB. The sensitivity and specificity of the network was 100% with a positive and negative predictive values of 100%. The reported artificial intelligence technology is inexpensive and can work with optical imaging systems currently used in most fertility clinics. Thus, the technology can be integrated into fertility clinics with very minimal principal and operational costs. Also, since our results suggest that this technology can perform with high accuracy reliably and rapidly, the reported technology in a clinical setting may significantly improve the clinical workflow for semen analysis.