Deep learning-enabled morphology analysis of bovine sperm for label-free imaging flow cytometry

Data analysis of sperm morphology is critical for evaluating bull fertility, yet it is commonly performed using light microscopy and staining techniques in a subjective and manual manner. In this study, we introduce a scalable, high-resolution approach combining label-free Imaging Flow Cytometry with deep learning for automated classification of bovine sperm morphology. We analyzed 401,535 single-cell images obtained out 1.8 million events acquired at 40 × magnification from three bull breeds - Kazakh Whitehead, Auliekol, and Simmental, from fresh and cryopreserved sperm - providing a uniquely large and diverse dataset. The dataset was used for training and evaluation of deep learning models. Multiple classification strategies and architectures were evaluated using consistent training and data preparation strategies, including MobileNetV3-Large, EfficientNetV2-S, ResNet-50, and ConvNeXt-Tiny. Among these, the ConvNeXt-Tiny yielded superior results, achieving an accuracy 91.1% and a macro F1-score of 0.91 after training with a Linear Probing and Fine-Tuning (LP-FT) strategy, and was chosen as the primary model to classify spermatozoa into eight distinct morphological categories. Testing across different conditions and breeds resulted in a 10-15% drop in generalization performance, highlighting the limitations in generalization due to dataset variability and underscoring the need for larger, standardized imaging protocols. The proportion of morphologically abnormal spermatozoa differed by seasons and following cryopreservation. This study highlights the advantages of integrating IFC and artificial intelligence algorithms for robust, high-throughput, and objective label-free assessment of spermatozoa morphology in both fresh and cryopreserved sperm, offering a promising tool for improving fertility diagnostics and breeding strategies in veterinary practice.