Studies on the assessment of ram sperm morphology

Abstract

Sperm morphology assessment is a widely used in vitro method for evaluating male fertility across livestock industries. Despite its prevalence, the test is hampered by considerable subjectivity and variation, seen among laboratories, assessment protocols, and individual morphologists, which undermines confidence in its reliability. This thesis investigated two major sources of variation in sperm morphology assessment: natural within-sire variation and human assessor bias. A longitudinal study of healthy rams demonstrated that sperm morphology can vary markedly across the spermatogenic cycle, reinforcing the need for multiple ejaculates to accurately characterise and evaluate an individual. Further investigation revealed significant disagreement among expert morphologists assessing the same sperm, highlighting the role of observer bias and the lack of standardised training. To address this, the thesis pursued a dual approach: (1) developing and validating a novel training tool for morphologists, and (2) trialling the use of machine learning to automate morphological classification. The training tool, built using expert-consensus-labelled images and incorporating principles of machine learning, significantly improved trainee accuracy across multiple classification systems. In contrast, machine learning models trained on the same dataset achieved only moderate accuracy, perhaps due to limitations in the amount of data available to train the model. Although promising, this experimental work demonstrated the practical challenges of developing and implementing machine learning at scale currently outweigh their benefits. Overall, this thesis demonstrates that improving standardised training for human observers offers a more efficient and immediately effective path toward improving the reliability of sperm morphology assessments. Nonetheless, continued development of objective tools remains an important future direction.