PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.14278v1cs.CVcs.AI

TL;DR

Sperm detection in microscopy is challenged by dense distributions, similar artifacts, and sperm-like impurities. SpermYOLO coordinates four adaptations in a compact YOLOv11-derived detector; it achieves strong joint SVIA results and remains effective in cross-scene SDTB evaluation.

  • Problem

    Dense distributions, visually similar artifacts, and sperm-like impurities make accurate sperm detection in microscopic images challenging.

    Source: §1 · p. 4

  • Method

    SpermYOLO is a compact YOLOv11-derived detector that coordinates feature discrimination, spatial–spectral semantic enhancement, adaptive multi-scale fusion, and detail-enhanced shared prediction.

    Source: §6 · p. 39

  • Results

    On SVIA, SpermYOLO achieves 97.2% sperm AP and 75.4% impurity AP, outperforming generic detectors, dedicated sperm models, and improved YOLO variants.

    Source: §6 · p. 39

  • Takeaways & Limitations

    SpermYOLO remains effective for sperm detection in testicular-biopsy microscopy with extremely small targets and complex tissue backgrounds.

    Source: §1 · p. 5, §6 · p. 39

    Evaluation covers semen-smear and testicular-biopsy microscopy; other sample-preparation and imaging conditions remain to be validated, and sperm tracking and motility estimation were not investigated.

    Source: §6 · p. 40

Abstract

from arXiv · show

Accurate sperm detection is essential for computer-assisted semen analysis, yet it remains challenging in microscopic images due to dense distributions, visually similar artifacts, and sperm-like impurities. In this paper, we propose SpermYOLO, a coordinated and compact YOLOv11-derived framework for joint sperm and impurity detection in microscopic images. SpermYOLO introduces four architectural improvements: C3k2-IDB for channel-wise discriminative feature extraction, D2SEM for spatial--spectral semantic enhancement, MFM for adaptive multi-scale feature fusion, and the DESD Head for detail-enhanced shared prediction. Experiments on the SVIA semen microscopic imaging benchmark show that SpermYOLO achieves 97.2\% sperm AP and 75.4\% impurity AP, outperforming generic detectors, dedicated sperm detection models, and improved YOLO variants. Compared with the baseline model, SpermYOLO improves sperm AP, impurity AP, $\mathrm{mAP}_{50}$, and $\mathrm{mAP}_{50:95}$ by 1.6, 10.0, 5.8, and 2.7 percentage points, respectively, while preserving a lightweight model scale. Cross-scene evaluation on the SDTB testicular-biopsy microscopy benchmark shows that SpermYOLO remains effective with extremely small sperm targets and complex tissue backgrounds, achieving the highest $\mathrm{mAP}_{50}$ and $\mathrm{mAP}_{50:95}$ of 74.8\% and 31.2\%, respectively. Ablation studies and qualitative analyses further support these improvements by demonstrating the contributions of the proposed modules and showing more focused feature response patterns than the baseline model. These findings suggest that SpermYOLO is an effective and efficient approach for sperm detection in challenging microscopic imaging scenarios.

1. Introduction

Sperm detection in microscopy remains challenging, especially with dense overlaps, similar artifacts, and sperm-like impurities. The introduction presents SpermYOLO as a compact, coordinated detector and reports strong joint detection and cross-scene results.

1. Introduction

  • Semen analysis is a cornerstone for clinical evaluation of male-factor infertility, while manual microscopic examination is time-consuming, labor-intensive, and subject to inter-operator variability.Sperm detection localizes cells as a basis for subsequent parameter quantification.Source: §1 · pp. 2–3
  • Dense distributions, overlaps, grayscale-similar artifacts, and sperm-like impurities make reliable localization and discrimination difficult in microscopy.Specialized models have improved detection, but robust joint sperm–impurity detection remains challenging under complex microscopic conditions.Source: §1 · p. 4
  • SpermYOLO coordinates four adaptations: C3k2-IDB for channel discrimination, D2SEM for spatial–spectral semantics, MFM for adaptive multi-scale fusion, and DESD for detail-enhanced shared prediction.The modules target feature extraction, semantic enhancement, fusion, and prediction, respectively.Source: §1 · pp. 4–5
  • On the SDTB testicular-biopsy benchmark, SpermYOLO remains effective with extremely small sperm targets and complex tissue backgrounds.This is a cross-scene evaluation beyond semen-smear microscopy.Source: §1 · pp. 5–6
  • Qualitative analyses compare SpermYOLO’s detection behavior and feature-response patterns with the baseline model.The introduction describes detection-result analyses and heatmaps.Source: §1 · p. 6
  • SpermYOLO achieves the highest joint sperm–impurity detection accuracy on SVIA among generic detectors, dedicated sperm models, and improved YOLO variants while maintaining a compact architecture.Figure 1 places it in the upper-right region for the two classes.Source: §1 · p. 5, Figure 1 · p. 6

2. Related Work

Prior sperm-detection work ranges from handcrafted image-processing methods to generic and specialized deep-learning detectors. Reported limitations include weak robustness to complex microscopy and unresolved difficulty distinguishing impurities from sperm.

2.1. Traditional Methods

  • Traditional threshold- and ellipse-based methods reported promising accuracy but relied on handcrafted priors, limiting robustness in dense overlaps and real-time applicability.The cited methods include threshold segmentation and morphological operations, filter-based enhancement, and elliptical detection.Source: §2.1 · p. 7

2.2. General Object Detection Models

  • Generic detectors demonstrated feasibility but showed low recall, artifact sensitivity, degradation in dense or low-quality images, and insufficient robustness in complex impurity-containing scenes.Reported issues include poor recognition of morphologically abnormal sperm and class-wise disparity favoring non-sperm structures.Source: §2.2 · pp. 7–8

2.3. Specialized Sperm Detection Models

  • Specialized models preserved small-object information using multi-scale fusion, cross-layer transfer, and attention, but blurred sperm, edge targets, and sperm-like impurities remained challenging.Spatial detail preservation alone does not fully resolve fine-grained discrimination.Source: §2.3 · pp. 8–9
  • Later specialized methods improved spatial-detail retention and efficiency, yet visually similar impurities remained insufficiently addressed.These approaches include depthwise convolution, attention, small-object heads, and overlapping image crops.Source: §2.3 · p. 9
  • Joint-detection studies targeted tiny-object features and scale variation, while SpermDet coordinated task-specific adaptations; balancing such design with efficiency remains important for CASA [43, 44, 45].SpermYOLO is presented as a compact one-stage detector with adaptations across feature extraction, semantic enhancement, fusion, and prediction.Source: §2.3 · pp. 9–10

3. Methodology

SpermYOLO adapts YOLOv11n across feature extraction, semantic refinement, fusion, and prediction to address low-contrast sperm and impurity cues while maintaining a compact design.

3. Methodology

  • YOLOv11n is the compact baseline, and SpermYOLO targets microscopy challenges across feature extraction, semantic enhancement, fusion, and prediction.The cited architecture overview frames the modifications as a coordinated adaptation rather than independent add-ons.Source: §3 · pp. 10–11, §3.1 · p. 11

3.2. C3k2-IDB: Inverted Discriminative Bottleneck

  • C3k2-IDB replaces C3k2 with an inverted, channel-recalibrated block designed to discriminate sperm and impurity cues with lightweight computation.The design retains full channel resolution during attention computation and avoids an additional channel bottleneck.Source: §3.2 · pp. 12–13

3.3. D2SEM: Dual-Domain Spectral Enhancement Module

  • D2SEM retains C2PSA self-attention while replacing its pointwise feed-forward network with SEFFN for joint spatial–spectral semantic refinement.The change introduces frequency-domain channel modulation for low-contrast sperm and impurity discrimination.Source: §3.3 · pp. 14–15

3.4. MFM: Multi-Scale Feature Fusion Module

  • MFM projects aligned neck inputs and uses pooled channel scores for scale-specific weights, replacing fixed concatenation with adaptive fusion.This is intended to reflect scale-dependent cues for sperm-head localization and discrimination from impurities and background clutter.Source: §3.4 · pp. 16–17

3.5. DESD Head: Detail-Enhanced Shared Detection Head

  • The DESD Head shares detail-enhanced processing and prediction projections across three detection scales, maintaining a compact structure while improving sensitivity to local details for localization.Its shared block uses detail-enhanced convolutions, followed by shared regression and classification projections.Source: §3.5 · pp. 17–19

4. Experimental Setup

The experiments use SVIA for joint sperm–impurity detection and SDTB for cross-scene sperm detection, with AP/mAP accuracy and model-efficiency measures. Training uses a shared 640 × 640 image size and dataset-specific epoch schedules.

4.1. Datasets

  • SVIA Subset-A contains 3,622 frames from 101 videos and over 125,000 annotated sperm and impurity objects; SDTB provides a patient-level cross-scene test with extremely small sperm targets.SDTB has 1,341 images and 5,548 sperm instances; 99.3% of instances are smaller than 16 × 16 pixels.Source: §4.1 · pp. 19–20

4.2. Evaluation Metrics

  • AP is the primary accuracy metric, with mAP50 and mAP50:95 reported alongside precision, recall, F1-score, and model-efficiency measures.Efficiency is measured by trainable parameters, GFLOPs, end-to-end FPS, and serialized weight size when applicable.Source: §4.2 · pp. 20–21

4.3. Implementation Details

  • Models were trained from scratch with SGD on 640 × 640 images, using 200 epochs for SVIA and 400 for SDTB, plus online augmentation.SpermYOLO hyperparameters were fixed across both datasets; augmentation included mosaic, horizontal flipping, HSV perturbation, scaling, and translation.Source: §4.3 · pp. 21–22

5. Results

The results section compares SpermYOLO with the baseline and examines proposed-module contributions on SVIA. It also covers comparisons with detector families, cross-scene testing on SDTB, and qualitative analysis.

  • The section compares SpermYOLO with the baseline and analyzes the contribution of each proposed module on SVIA.Source: §5 · p. 22
  • It compares SpermYOLO with representative generic, sperm-specific, and improved YOLO detectors.Source: §5 · p. 22
  • The evaluation also includes cross-scene testing on SDTB and qualitative analysis.Source: §5 · p. 22

5.1. Baseline Comparison and Ablation Study

SpermYOLO improves on the baseline while retaining a compact model, and the ablations show that module combinations and placement affect the balance of detection metrics. The study also selects 256 DESD hidden channels and places C3k2-IDB in both backbone and neck.

5.1. Baseline Comparison and Ablation Study

  • Over the 200-epoch schedule, both models’ training losses converge, while SpermYOLO later attains higher validation recall, mAP50, and mAP50:95 with validation precision close to baseline.The test checkpoint for each model was selected using its best validation performance.Source: §5.1 · pp. 22–23, Figure 7 · p. 23

5.1.1. Module Ablation and P-R Analysis

  • Among single-module variants, DESD raises API from 65.4% to 74.2% and mAP50 from 80.5% to 85.1%, while MFM raises recall from 73.8% to 78.9%.D2SEM raises API by 4.4 points and has the highest single-module precision at 87.1%; C3k2-IDB raises mAP50 by 2.2 points.Source: §5.1.1 · pp. 23–24
  • SpermYOLO raises mAP50:95 from 44.5% to 47.2%, retains 2.6M parameters, reduces GFLOPs from 6.3 to 5.9, and improves recall from 73.8% to 81.6%.Its F1-score rises from 78.7% to 82.7%; relative to the strongest three-module combination, it has slightly lower mAP50 and mAP50:95 but higher APS, precision, and F1-score.Source: Table 1 · p. 24, §5.1.1 · p. 26
  • Two-module combinations show trade-offs: C3k2-IDB+DESD Head reaches 82.0% recall and 80.7% F1-score, while D2SEM+DESD Head has the highest two-factor APS (97.0%) and precision (88.5%).MFM+DESD Head reaches mAP50 of 81.7% and mAP50:95 of 43.9%.Source: §5.1.1 · pp. 24–25
  • Among three-module variants, C3k2-IDB+D2SEM+DESD Head has the highest mAP50 (87.3%) and mAP50:95 (47.6%), while its precision falls to 79.5%.Without DESD Head, C3k2-IDB+D2SEM+MFM has lower recall and aggregate AP; without C3k2-IDB, D2SEM+MFM+DESD Head attains API of 77.6%.Source: §5.1.1 · pp. 25–26
  • At the F1-maximizing confidence threshold, SpermYOLO precision is 84.2% versus the baseline’s 85.5%, yet its curve retains higher precision across most medium- and high-recall ranges.The curve-level comparison is consistent with the AP-based metrics.Source: §5.1.1 · p. 26, Figure 8 · p. 27

5.1.2. Ablation Experiments for DESD Hidden Channels

  • The 256-hidden-channel DESD Head has the best tested overall performance; widening it to 512 raises GFLOPs from 5.9 to 9.8 while mAP50 falls to 85.6%.At 1024 channels, parameters and GFLOPs rise to 3.8M and 25.0 while mAP50 falls to 84.8%; 128 channels has the lowest compute but mAP50:95 of 44.8% and F1-score of 78.1%.Source: §5.1.2 · pp. 27–28, Table 2 · p. 28

5.1.3. Ablation Experiments for C3k2-IDB Placement

  • Placing C3k2-IDB in both backbone and neck gives the best tested overall performance and slightly lower GFLOPs than either single-placement variant.Backbone-only and neck-only settings have similar mAP50 values of 83.3% and 83.6%, respectively; neck-only has higher API and mAP50:95 but lower precision and F1-score.Source: §5.1.3 · pp. 28–29, Table 3 · p. 29

5.2. Comparison with Representative Detectors on SVIA

On SVIA, SpermYOLO leads the evaluated detectors on both sperm and impurity AP, as well as aggregate AP metrics. It also retains a compact computational profile and a balanced precision–recall trade-off.

5.2. Comparison with Representative Detectors on SVIA

  • SpermYOLO achieves the highest AP for both classes on SVIA: 97.2% for sperm, 0.4 points above ACTIVE, and 75.4% for impurities, 2.4 points above YOLOv5n-SPD.Several competing methods have high sperm AP but lower impurity AP, including ACTIVE (96.8% APS versus 62.7% API).Source: Table 4 · p. 30, §5.2 · pp. 30–31
  • SpermYOLO has the highest mAP50, mAP50:95, and F1-score among evaluated detectors, exceeding YOLOv5n-SPD by 2.0 and 5.6 points on the two mAP metrics.Compared with SpermDet, the strongest sperm-specific model in mAP50, its mAP50 and mAP50:95 gains are 3.9 and 2.3 points.Source: Table 4 · p. 30, §5.2 · p. 31
  • Although MAF-YOLO has the highest recall (85.1%), its precision is 71.6% and mAP50 is 79.2%; SpermYOLO has the highest F1-score among evaluated detectors.YOLOv11n, YOLOv12n, and YOLOv5n-SPD have slightly higher precision, while MAF-YOLO and SpermDet have higher recall.Source: §5.2 · p. 31
  • At 5.9 GFLOPs, SpermYOLO has the lowest computational cost among compared models, with 2.6M parameters and 5.7MB weights.Its 185 FPS is below MAF-YOLO’s 220 and CST-YOLO’s 191, but the paper reports it as sufficient for real-time automated analysis.Source: §5.2 · pp. 31–32

5.3. Cross-Scene Evaluation on SDTB

On SDTB, a cross-scene test using testicular-biopsy images with sperm annotations only, SpermYOLO leads the evaluated detectors on mAP50 and mAP50:95. It also has the highest precision and F1-score despite slightly lower recall than ACTIVE.

5.3. Cross-Scene Evaluation on SDTB

  • Several general-purpose detectors have low SDTB mAP50, including Faster R-CNN at 14.4%, SSD at 12.4%, and RT-DETR-l at 37.3%.YOLOv5n-SPD, despite competitive SVIA performance, ranks second-lowest among evaluated YOLO models on SDTB.Source: §5.3 · p. 32
  • Against ACTIVE, the strongest competitor in mAP50, SpermYOLO improves mAP50 by 0.8 points and mAP50:95 by 4.2 points on SDTB.The paper relates the stronger mAP50 to target identification and the higher mAP50:95 to fine-grained localization under tiny-target and complex-background conditions.Source: §5.3 · p. 33
  • On SDTB, SpermYOLO records the highest precision (76.3%) and F1-score (71.4%), while its recall of 67.1% is slightly below ACTIVE’s 67.8%.The paper concludes that its leading AP metrics and precision–recall balance support effectiveness in this distinct microscopy setting.Source: §5.3 · p. 33
  • SpermYOLO achieves the highest SDTB mAP50 (74.8%) and mAP50:95 (31.2%) among evaluated detectors, exceeding YOLOv8n by 3.0 and 1.3 points, respectively.The dataset differs from SVIA in imaging source and annotation scope, testing transfer to testicular-biopsy microscopy.Source: §5.3 · p. 32, Table 5 · p. 34

5.4. Qualitative Analysis

Qualitative comparisons show fewer false positives and missed targets with SpermYOLO in representative SVIA scenes, alongside more localized heatmap responses than the baseline. The analyses cover detection behavior and spatial feature responses.

5.4.1. Detection Result Visualization

  • In representative SVIA scenes, SpermYOLO reduces false positives from sperm-like background structures and recovers more small or low-contrast targets than the baseline.In a crowded sperm-dominant example, the baseline has one false positive and misses two sperm, while SpermYOLO has no obvious false positives or missed detections.Source: Figure 9 · p. 35, §5.4.1 · pp. 35–36
  • Figure 9 places baseline predictions in the upper row and SpermYOLO predictions below; green, red, and blue boxes mark correct detections, false positives, and missed targets.The examples include sparse scenes with sperm-like particles and a dense mixed scene with background interference.Source: §5.4.1 · pp. 34–35

5.4.2. Attention Heatmap Analysis

  • Grad-CAM++ heatmaps show more compact, selective SpermYOLO activations around salient microscopic structures, with fewer scattered and diffuse background responses than the baseline.The baseline shows point-like and diffuse responses over noise and non-target regions, while SpermYOLO retains more localized responses.Source: §5.4.2 · pp. 36–37, Figure 10 · p. 38
  • Figure 10 rows show the original image, baseline heatmap, and SpermYOLO heatmap from left to right; warmer colors indicate stronger response.The heatmaps combine upsampled responses from final feature maps before prediction at multiple detection scales.Source: §5.4.2 · pp. 36, 38

6. Conclusion

SpermYOLO combines four coordinated adaptations for joint sperm–impurity detection and performs strongly on semen-smear and testicular-biopsy microscopy benchmarks. Its evaluation remains limited to these two imaging scenarios, and sperm tracking and motility analysis were not investigated.

6. Conclusion

  • C3k2-IDB, D2SEM, MFM, and the DESD Head target channel discrimination, deep semantic enhancement, adaptive multi-scale fusion, and detail-sensitive shared prediction, respectively.Together, these adaptations address small, low-contrast sperm targets and sperm-like impurities.Source: §6 · p. 39
  • On SVIA, SpermYOLO achieves 97.2% sperm AP and 75.4% impurity AP, outperforming generic detectors, dedicated sperm detection models, and improved YOLO variants.Experiments on SVIA and SDTB are reported as demonstrating the framework’s effectiveness.Source: §6 · p. 39
  • On SDTB, SpermYOLO remains effective for sperm detection amid extremely small targets and complex tissue backgrounds.This is the paper’s cross-scene evaluation in testicular-biopsy microscopy.Source: §6 · p. 39
  • Qualitative analyses report improved detection behavior and more focused feature response patterns.The passage does not specify a quantitative measure for these qualitative observations.Source: §6 · p. 40
  • Evaluation covers semen-smear and testicular-biopsy microscopy, while other preparation and imaging conditions and sperm tracking and motility analysis remain uninvestigated.Future work proposes validation under more diverse microscopy conditions and integration with tracking and motility analysis.Source: §6 · p. 40

CRediT authorship contribution statement

The statement assigns authorship contributions across conceptualization, methodology, software, investigation, validation, writing, visualization, resources, data curation, funding, project administration, and supervision.

CRediT authorship contribution statement

Loading 2609.14278v1…