149. Segmentation guided deep learning for sperm morphology classification in bull breeding soundness examinations.

Abstract Breeding soundness examination (BSE) is one of the most critical management practices to ensure the profitability of a beef herd. Semen analysis of sperm motility and morphology is time-consuming, requires significant training, and is susceptible to human error and subjectivity. The objective of this study was to evaluate the use of computer vision for sperm morphology classification in bulls. We hypothesized that computer vision can accurately perform sperm morphology evaluations compared with human-based classifications. Bos taurus bulls were exposed to BSE following the guidelines of the Society for Theriogenology. Morphology slides were prepared chute-side using eosin-nigrosin staining. Slides were examined using bright-field microscopy at 1000x magnification, and images were captured. Individual sperm cells were annotated using CVAT to segment whole sperm and regions of the sperm (head, midpiece, tail). Individual sperm images were labelled by three trained evaluators as morphologically normal or by location of abnormality. Cell images (n = 473) were used to train a binary baseline convolutional neural network (CNN) model (normal vs. abnormal), and two multi-class CNN models were developed to classify five categories based on defect localization (morphologically normal, head, midpiece, tail, or other defect). The Multiclass Full Image model received the entire segmented sperm as input and predicted the single most biologically relevant defect location. The Multiclass Regional model used anatomical segmentation masks to isolate the head, midpiece, and tail as separate crops, training one independent binary classifier per region, allowing a single cell to receive multiple simultaneous defect classifications. Both models used transfer learning with ImageNet-pretrained backbones, nested cross-validation for hyperparameter selection, and consistent data augmentation and pixel normalization. Performance was assessed using accuracy, macro precision, recall, macro F1-score, and balanced accuracy, with regional model metrics reported independently per anatomical region. The binary baseline model achieved a balanced accuracy of 0.794 and macro F1-score of 0.792, partially distinguishing normal versus abnormal sperm cells. Among the multiclass models, the Multiclass Full Image model reached a balanced accuracy of 0.559 and macro F1-score of 0.537. The Multiclass Regional model improved on the ability to predict defect location with balanced accuracies of 0.664, 0.816, and 0.822 and macro F1-scores of 0.705, 0.831, and 0.841 for head, midpiece, and tail regions, respectively. Within this preliminary dataset, the Multiclass Regional model outperformed the Multiclass Full Image model in predicting primary defect location while retaining the ability to identify concurrent secondary defects, indicating that anatomical region-specific classification better reflected the complexity of sperm morphology abnormalities. While classification by defect location shows promising trends, further improvements are required and will likely depend on increased dataset size and more balanced class representation.