Domain-Aware and Multi-Scale Deep Learning Framework for Sperm Morphology Classification

Automated sperm morphology classification can reduce the subjectivity and workload of manual microscopic assessment, but limited benchmark data and subtle morphology-related visual differences remain important challenges. This study proposes a domain-specific ConvNeXt-Tiny-based framework for sperm morphology classification on the SMIDS benchmark. The framework combines Hi-LabSpermMorpho-based domain-specific pretraining, multi-scale feature extraction, attention-guided feature refinement, class-balanced sampling, test-time augmentation, MixUp, label smoothing, focal-loss-based optimization, and supervised contrastive learning. Domain-specific pretraining was compared with ImageNet initialization, and the framework was examined through cumulative ablation, component-removal experiments, and matched-condition architecture comparisons. The best configuration, combining Hi-LabSpermMorpho pretraining with 10× augmented training data, achieved a patch-level average accuracy of 92.73%, numerically the highest among the evaluated configurations and backbones. In matched-condition architecture comparisons, the differences between the proposed framework and EfficientNetV2-S, Swin-T, and domain-pretrained BEiT-Base were not statistically significant. The full training configuration significantly outperformed the baseline configuration under 5× augmentation, whereas the component-removal experiments did not establish the necessity of each individual component. Because the held-out test folds were also used for checkpoint selection, the reported accuracies are subject to selection bias and cannot be regarded as independent test estimates; accordingly, all comparisons should be regarded as exploratory. These findings suggest that domain-specific pretraining can support sperm morphology classification, although independent validation is required.