SpermYOLO: A Coordinated YOLO-Based Detector for Accurate and Efficient Sperm and Impurity Detection in Microscopic Images
Shengqi Chen, Zilin Wang, Xingyu Pan, Wenting Yu, Pengchao Deng, Guohua Wu
cs.CV · cs.AI
Quick verdict
YOLOv11n-derived SpermYOLO lifts SVIA sperm/impurity AP to 97.2/75.4; ablations are thorough, but single-run margins are small and novelty is combinational.
Summary
The paper presents SpermYOLO, a YOLOv11n-derived detector for joint sperm and impurity detection in semen microscopy, modifying four pipeline stages: C3k2 blocks become inverted bottlenecks with EffectiveSE (C3k2-IDB), C2PSA gains an FFT-based spectral feed-forward network from TransMamba (D2SEM), neck concatenation is replaced by MLP-softmax channel-weighted fusion (MFM), and the head uses detail-enhanced convolution with cross-scale parameter sharing (DESD Head). On SVIA Subset-A (2,146/659/817 images, video-level split), it reports 97.2% sperm AP, 75.4% impurity AP, 86.3% mAP50 and 47.2% mAP50:95 versus 95.6/65.4/80.5/44.5 for the baseline — gains of 1.6/10.0/5.8/2.7 points — at 2.6M parameters and 5.9 GFLOPs. It tops a 17-model comparison including ACTIVE and SpermDet, though margins over the strongest rivals are just +0.4 sperm AP and +2.4 impurity AP. On SDTB testicular-biopsy images it achieves the best mAP50 (74.8%) and mAP50:95 (31.2%). Ablations cover all module combinations plus hidden-channel and placement sensitivity, but standalone MFM lowers mAP50:95 below baseline. All results are single runs without significance testing, the modules are largely adapted from prior works, and qualitative conclusions rest on figures not verifiable from the text. Code and public datasets are stated to be available.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Four coordinated modifications to YOLOv11n (C3k2-IDB, D2SEM, MFM, DESD Head) jointly improve joint sperm-impurity detection over the baseline on SVIA Subset-A by 1.6/10.0/5.8/2.7 points on sperm AP/impurity AP/mAP50/mAP50:95.
- On SVIA Subset-A, SpermYOLO attains the best joint accuracy among 17 compared detectors (97.2% sperm AP, 75.4% impurity AP) while keeping 2.6M parameters and 5.9 GFLOPs.
- The design transfers across scenes: on SDTB testicular-biopsy images with ~10-pixel sperm it achieves the best mAP50 (74.8%) and mAP50:95 (31.2%) among evaluated detectors.
- Ablations (Table 1) support a positive contribution of each proposed module in combination.
Strengths
Unusually complete ablation design
Table 1 covers all single-, two-, and three-module combinations (16 configurations), plus a hidden-channel sweep (Table 2, 128-1024) and a C3k2-IDB placement study (Table 3, backbone/neck/both) — more thorough than typical in this genre.
Trade-offs disclosed rather than hidden
The precision drop vs baseline (84.2% vs 85.5%) is explained via P-R curves (Fig. 8), and the full model's mAP50 below the best three-module variant (86.3% vs 87.3%) is openly discussed with a sperm-priority rationale.
Leakage-aware evaluation protocol
Efficiency claims backed by numbers
Cross-benchmark validation at very different target scales
Weaknesses
Combinational novelty only
No statistical uncertainty
MFM's individual value is questionable
Threshold-dependent P/R/F1 reporting
Qualitative and heatmap claims not independently checkable
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SpermYOLO (arXiv:2609.14278) is a YOLOv11n-derived detector for joint sperm and impurity detection in CASA microscopy. The authors claim four coordinated improvements — C3k2-IDB (inverted bottleneck + EffectiveSE), D2SEM (FFT-spectral feed-forward inside partial self-attention), MFM (MLP-softmax weighted neck fusion), and the DESD Head (detail-enhanced convolution with cross-scale parameter sharing) — that lift SVIA Subset-A results from 95.6/65.4 sperm/impurity AP to 97.2/75.4 and mAP50 from 80.5% to 86.3%, lead a 17-model comparison, and transfer to SDTB (74.8% mAP50, 31.2% mAP50:95) where sperm average only ~9.9 pixels. The quantitative claims are internally consistent and mostly supported by the presented tables, with caveats about run variance, baseline comparability, and configuration selection.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does SpermYOLO achieve for sperm detection?
On the SVIA Subset-A benchmark the paper reports 97.2% sperm AP and 75.4% impurity AP, with 86.3% mAP50 and 47.2% mAP50:95, improving the YOLOv11n baseline by 1.6/10.0/5.8/2.7 points respectively at 2.6M parameters and 5.9 GFLOPs. On the SDTB testicular-biopsy set it reaches the best mAP50 (74.8%) and mAP50:95 (31.2%) among the compared detectors.
Is SpermYOLO code and data available?
Yes. The paper points to a public GitHub repository (github.com/ElvisChen7/SpermYOLO), and both benchmarks (SVIA, SDTB) are public datasets with fixed video-level and patient-level splits. Training and evaluation are therefore repeatable in principle; full retraining needs a GPU (experiments used an RTX 3090, 200-400 epochs).
Is SpermYOLO real-time and suitable for CASA systems?
The paper reports 185 FPS end-to-end on an RTX 3090 (batch size 16) at 5.9 GFLOPs with 2.6M parameters, which the authors argue is sufficient for real-time CASA front-ends. Note that the study evaluates detection only, not sperm tracking or motility estimation.
What is new in SpermYOLO compared with plain YOLOv11?
Four swaps: C3k2 blocks become MBConv+EffectiveSE inverted bottlenecks (C3k2-IDB), C2PSA gains an FFT-based spectral feed-forward network (D2SEM), neck concatenation becomes MLP-softmax weighted fusion (MFM), and the head uses detail-enhanced convolution with cross-scale parameter sharing (DESD Head). Each component adapts prior techniques (EfficientNet, CenterMask, TransMamba, DCMP-Net, DEA-Net); the contribution is their coordinated use for sperm microscopy.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-10 (scatter plot, detection examples, Grad-CAM++ heatmaps) appear only as caption text in the extraction; all figure-dependent qualitative claims could not be assessed.
- Text extraction is otherwise complete through the reference list; all cited numbers were cross-checked against Tables 1-5 and the abstract.
- arXiv v1 (13 Sep 2026); no peer-review venue or status information is available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.