HEIR-Nets: Hierarchical Ensemble with Inheritance-Based Refinement Networks for Sperm Target Detection with Noisy Labels

Sperm motility assessment in Assisted Reproductive Technology critically relies on accurate detection of biomedical targets within electron microscopic sperm images. However, annotating numerous small targets per image poses significant challenges, inevitably introducing noisy labels. These label errors severely degrade the learning efficacy of existing deep architectures, limiting high-precision sperm detection and classification. Therefore, we propose the Hierarchical Ensemble with Inheritance-based Refinement Networks (HEIR-Nets) featuring three synergistic strategies: 1) Self-Correcting Pseudo-label Inheritance (SPI) performs dual-action correction by analyzing pseudo-label confidence from the current branch model, while incorporating random target dropout intervention to generate self-corrected sample sets for inheritance by subsequent branches. 2) Multi-model Knowledge Inheritance (M2KI) initializes current branch models using predecessor branch model parameters, while iteratively training them on inherited purified datasets. 3) Multi-branch Fusion (MBF) integrates dual-weighted decisions by combining designed gamma-encoded model credibility weights with sample confidence weights. The performance of HEIR-Nets is verified on the self-built sperm image dataset with noisy labels of different proportions and the public sperm image dataset. The results validate HEIR-Nets’ superiority and robustness. In particular, on the public dataset with 20% noisy labels, the F1 score and accuracy of HEIR-Nets exceeded those of the state-of-the-art methods by 4.56% and 7.54%, respectively.