P-075 Artificial intelligence for sperm analysis: A KNIME-Based automated analysis of sperm morphology

Abstract Study question To evaluate an automated sperm morphology assessment method using the KNIME Analytics Platform compared to manual analysis performed as per WHO guidelines. Summary answer Our emphasizes how artificial intelligence technologies have potential foster standardization of with comparable precision and reliability. What is known already Manual considered most difficult parameter standardize due its subjective nature, strongly linked operator's level expertise. Indeed, there a high degree inter intra-laboratory variability. examination time-consuming laborious. There were many attempts automate analysis, especially CASA (Computer Assisted Sperm Analysis) systems, but their performance still disputable. One difficulties in this field study lack publicly available datasets. Besides, databases are only focused on head morphology. design, size, duration A total 37 semen samples from men attending our laboratory for infertility investigation included, over period one year. For each sample, smears fixed stained by Spermoscan® kit. Participants/materials, setting, methods 1000 images individual spermatozoa obtained MMC® system. The number sample depended quality. Three experts classified these according modified David classification results then processed algorithm created teh Platform, trained tested classify spermatozoa. This workflow uses CNN (convolutional neural network) perform image dataset. Main role chance Of analyzed, we counted : 116 Normal morphology, 67 abnormal post-acrosomal region, 128 acrosomes, 8 elongated heads, 6 thin heads,10 microcephalic,7 multiple 27 coiled tails, 7 cytoplasmic residues, 17 angulated short 4 697 associated abnormalities. image, notebook file containing abnormalities assessed three was created, addition head, mid-piece tail dimensions dataset randomly partitioned into 2 groups: 80% data formed training set, remaining 20% fully independent test set. best KNIME-Based achieved (97 % True positive rate), worst (69 rate). machine learning model classifies at accuracy (99.5 %). overall process occurs less than 10 seconds. Limitations, reasons caution database included various midpiece anomalies. However, category unequal limited occurrence some morphological Thus, it important increase categories obtain better results. Wider implications findings goal expand David’s classification-based database. We aim improve model, put service routine use laboratories. Trial registration not applicable