A generalized classification of human sperm head morphology via Contrastive Meta-learning with Auxiliary Tasks
Yuh‐Shyan Chen, Yu-Chi Chang, Jiayi Hong · Neurocomputing · 2025
Semen analysis is the primary method for evaluating male infertility, with sperm morphology being a critical indicator. However, existing methods for classifying human sperm head morphology (HSHM) often lack cross-domain generalizability. To address this limitation, we propose an enhanced meta-learning algorithm that learns invariant features across tasks, improving generalization by transferring knowledge to new tasks. To mitigate gradient conflicts in multi-task learning, we separate meta-training tasks into primary and auxiliary tasks. This approach, in conjunction with auxiliary tasks, enhances the model’s generalization using diverse HSHM datasets. We introduce the Contrastive Meta-Learning with Auxiliary Tasks (HSHM-CMA) algorithm, which integrates localized contrastive learning in the outer loop of meta-learning to exploit invariant sperm morphology features across domains. This improves task convergence and adaptation to new categories. In our evaluation, we assess the model’s generalization performance using three testing objectives: the same dataset with different HSHM categories, different datasets with the same HSHM categories, and different datasets with different HSHM categories. Our evaluation across these objectives demonstrates that HSHM-CMA outperforms existing meta-learning approaches, achieving accuracies of 65.83%, 81.42%, and 60.13%, respectively.