SpeHeaTal: A Cluster-Enhanced Segmentation Method for Sperm Morphology Analysis

Yi Shi, Yun-Kai Wang, Xu-Peng Tian, Tie-Yi Zhang, Bing Yao, Hui Wang, Yong Shao, Cen-Cen Wang, Rong ZengAAAI 2025 · DOI

Introduction

Analyzing sperm morphology—the shape and structure of sperm—is a cornerstone of male fertility diagnostics. However, manual assessment is slow, subjective, and labor-intensive for clinicians. While deep learning offers a potential solution for automation, it faces three major hurdles in clinical practice: the scarcity of expert-annotated datasets, the frequent occurrence of overlapping sperm structures, and the presence of dye impurities that confuse automated systems. Existing models often fail to resolve these overlaps or require extensive labeled data, making them impractical for real-world use.

Methodology: The SpeHeaTal Framework

To address these challenges, we present SpeHeaTal (Sperm Head and Tail), an unsupervised framework that segments sperm in clinical microscopic images without requiring massive labeled datasets. The system follows a "decomposition-combination" pipeline.

Head Segmentation

The system utilizes the Segment Anything Model (SAM) to identify potential objects. To distinguish actual sperm heads from dye impurities or debris, two specialized filters are applied: a Shape Index (SI) to identify oval structures and a Color Index (CI) to detect the characteristic purple hue of dyed sperm heads.

Tail Segmentation (CON2DIS)

The core innovation of SpeHeaTal is the CON2DIS clustering algorithm, designed specifically to untangle overlapping tails. Traditional distance-based clustering often fails at intersections because crossing tails are physically close. CON2DIS resolves this by evaluating three factors: Connectivity (pixel continuity), Conformity (alignment of tangent angles), and Distance. By reducing tails to one-pixel-wide skeletons, the algorithm can logically trace a tail's path through a crossover point.

Illustration of the CON2DIS algorithm resolving a tail intersection using connectivity and conformity principles.

Mask Splicing

In the final stage, a tailored splicing technique matches the segmented heads to their respective tails based on the distance between endpoints and the similarity of their slopes.

Experimental Evaluation

The researchers collaborated with several hospitals to create a realistic clinical dataset comprising ~2,000 unlabeled images for calibration and 240 expert-annotated images for testing. Unlike "clean" public datasets, this collection specifically includes the dense overlaps and impurities found in actual clinical practice.

SpeHeaTal significantly outperformed both supervised and unsupervised baselines. It achieved a mean Intersection over Union (mIoU) of 72.21 and a mean Dice (mDice) of 81.07. In comparison, the supervised U-Net scored 49.21 mIoU, while the standard unsupervised SAM scored only 34.17 mIoU.

SpeHeaTal segmentation results demonstrating successful separation of overlapping sperm tails.

Discussion and Conclusion

SpeHeaTal demonstrates that foundation models like SAM can be effectively adapted for specialized biological tasks when combined with geometric algorithms like CON2DIS. By removing the need for labeled training data, this method is easily deployable in new clinical settings.

While the algorithm reaches a "breaking point" in extremely complex tangles (more than ten overlapping sperm), it remains robust for standard clinical density. Future work will focus on refining the splicing logic for high-density environments and potentially integrating temporal data from sperm videos to further enhance diagnostic accuracy.

Summary updated

BibTeX
@inproceedings{shi2025speheatal,
  author = {Yi Shi and Yun-Kai Wang and Xu-Peng Tian and Tie-Yi Zhang and Bing Yao and Hui Wang and Yong Shao and Cen-Cen Wang and Rong Zeng},
  title = {{SpeHeaTal: A Cluster-Enhanced Segmentation Method for Sperm Morphology Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i1.32055},
}
Figure