Research map / Detection and segmentation / Open-set and few-shot detection
SpeHeaTal: A Cluster-Enhanced Segmentation Method for Sperm Morphology Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SpeHeaTal, an unsupervised sperm segmentation pipeline that uses SAM for heads and dye filtering, a new connectivity-conformity-distance clustering method (ConConDis) for overlapping tails, and mask splicing to join heads and tails.
Problem
Segmenting individual sperm in clinical microscopy images without labels is hard due to overlapping sperm tails and dye impurities; existing supervised and unsupervised methods perform poorly on these.
Approach
A clustering algorithm for overlapping tails that jointly combines distance (kNN), tangent-space conformity (principal angles from mixtures of PPCA, as in SMMC), and connectivity (DBSCAN-filtered connected components) in a spectral framework, combined with a decomposition-combination pipeline in which SAM handles heads and tails are spliced back using endpoint distance and slope similarity.
Key results
- SpeHeaTal achieves 72.21 mIoU and 81.07 mDice on the in-house sperm dataset, versus 49.21/62.17 for U-Net and 34.17/46.75 for SAM.
- In the clustering ablation, the proposed ConConDis-style clustering reaches 72.21 mIoU vs. 46.86 for spectral clustering and 14.63 for K-Means when used for tail segmentation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpeHeaTal instead of U-Net | Clinical sperm segmentation · Clinical sperm dataset | mIOU: 49.21 -> 72.21 | Helped |
| Use SpeHeaTal instead of U-Net | Clinical sperm segmentation · Clinical sperm dataset | mDice: 62.17 -> 81.07 | Helped |
| Replace C3 clustering with K-Means (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mIOU: 14.63 -> 72.21 | Hurt |
| Replace C3 clustering with K-Means (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mDice: 26.94 -> 81.07 | Hurt |
| Replace C3 clustering with AHC (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mIOU: 27.76 -> 72.21 | Hurt |
| Replace C3 clustering with AHC (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mDice: 38.17 -> 81.07 | Hurt |
| Replace C3 clustering with LSA (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mIOU: 37.61 -> 72.21 | Hurt |
| Replace C3 clustering with LSA (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mDice: 45.92 -> 81.07 | Hurt |
| Replace C3 clustering with SC (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mIOU: 46.86 -> 72.21 | Hurt |
| Replace C3 clustering with SC (ablation) | SpeHeaTal tail segmentation · Clinical sperm dataset | mDice: 56.33 -> 81.07 | Hurt |
| Increase the distance and angle thresholds in mask splicing (ablation) | SpeHeaTal · Clinical sperm dataset | mDice | Hurt |
Surprising findings
- Unsupervised SAM-based and U2Seg/CutLER baselines perform worse than supervised U-Net on this dataset, with U2Seg reaching only 16.44 mIoU.
Limitations
- Struggles when more than about ten sperm are heavily entangled.
- Requires adequate dilution of semen samples in clinical practice.
- Evaluated on a single in-house dataset of 240 annotated images.
Experimental setup
Datasets: clinical sperm image dataset (~2,000 unlabeled images) expert-annotated clinical sperm images (240)
Benchmarks: in-house clinical sperm dataset (240 expert-annotated images)
Metrics: mIoU mDice
Baselines: U-Net IUHS CN2UA SAM U2Seg CutLER K-Means AHC LSA SC
Builds on: Segment Anything Model Spectral Multi-Manifold Clustering DBSCAN CutLER
Closest papers by meaning
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays (NeurIPS 2026 · similarity 0.78)
- GraphMorph: Tubular Structure Extraction by Morphing Predicted Graphs (NeurIPS 2024 · similarity 0.77)
- MultiOrg: A Multi-rater Organoid-detection Dataset (NeurIPS 2024 · similarity 0.77)
- Mudslide: A Universal Nuclear Instance Segmentation Method (CVPR 2024 · similarity 0.77)
- SCCS: Deep Neural Spectral Clustering for Self-Supervised Subcellular Structure Segmentation (AAAI 2025 · similarity 0.77)
- OnlineAnySeg: Online Zero-Shot 3D Segmentation by Visual Foundation Model Guided 2D Mask Merging (CVPR 2025 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.