Integrated subcellular localization of functional fluorescence probes and functional analysis in motile spermatozoa by an AI-enhanced algorithm

Abstract

In the evaluation of male infertility, precise assessment of sperm functional competence has surpassed the requirements of conventional semen parameters. Existing computer-aided analysis systems are deficient at the molecular diagnostic level and also face challenges in live-cell fluorescence quantification. To address these issues, we have developed a novel integrated computational-imaging platform that combines a fine-tuned You Only Look Once version 8 (YOLOv8) architecture, tailored for the EVISEN dataset, with dual-probe fluorescence microscopy image segmentation, enabling simultaneous quantification of intracellular pH (pHi) and mitochondrial DNA G-quadruplexes (mtDNA G4s). By automating the localization of fluorescent foci, our algorithm systematically discriminates between the fluorescent signatures of the sperm head and principal piece, revealing correlations between fluorescence intensity ratios and sperm functional outcomes. This study demonstrates the potential of artificial intelligence (AI)-enhanced multimodal sperm analysis for molecular phenotyping of sperm functional competence. Integrating deep learning with live-cell fluorescence imaging, our platform offers a transformative tool for mechanistically informed diagnostics of male infertility.

INTRODUCTION

Infertility, as defined by the World Health Organization (WHO), constitutes a failure to achieve clinical pregnancy following ≥12 months of regular unprotected sexual intercourse.1 This reproductive health challenge affects an estimated 15% of couples globally,2 with male factors contributing to 30%–50% of cases, which can be manifested as impairments in spermatogenesis and semen parameter abnormalities (encompassing morphological defects, motility deficiencies, oligozoospermia, etc.).3,4 In vitro fertilization (IVF) has emerged as a cornerstone intervention for addressing infertility, necessitating rigorous sperm quality assessment and selection optimization.5

Conventional semen analysis relies on manual microscopic evaluation by skilled technicians, a methodology limited by substantial intra- and inter-observer variability.6 This inherent subjectivity underscores the imperative for standardized analytical approaches. Computer-aided sperm analysis (CASA) systems address these limitations through automated quantification of sperm concentration, motility parameters, and morphological characteristics.7,8 While CASA provides improved objectivity in semen parameter assessment,9 its diagnostic utility remains constrained by two critical gaps: (1) limited correlation between conventional semen parameters and actual fertility outcomes10 and (2) inability to elucidate etiological mechanisms underlying abnormal sperm characteristics.11

Our previous research has identified two seminal biomarkers of sperm functional competence: intracellular pH (pHi) dynamics and mitochondrial DNA G-quadruplex structures (mtDNA G4s).12,13 The selection of pHi and mtDNA G4s as biomarkers was based on their complementary mechanistic rationale and clinical detectability. pHi is a regulator of sperm capacitation through CatSper channel modulation, with established correlations between pathological alkalization and reduced IVF success rates. While pHi assessment benefits from standardization from 2’,7’-Bis(2-carboxyethyl)-5(6)-carboxyfluorescein (BCECF) probe measurements, mtDNA G4s offer novel insights into oxidative phosphorylation efficiency, with preliminary data demonstrating that G4 destabilization is correlated with mitochondrial dysfunction. Accumulating research has shown that mtDNA G4s are dynamically regulated and provide valuable insights into metabolic reprograming in stress condition, used in metabolic evaluation of cancer for therapeutical purposes.14 We have designed novel mtDNA G4s probe and have shown that in flow cytometry, they can effectively monitor etiological biomarkers in patients with abnormal fertility; however, whether these probes can be co-monitors in live imaging in motile spermatozoa is unknown.

In the current study, we propose a dual-parameter strategy that synergizes real-time functional assessment (pHi dynamics) with structural–metabolic evaluation (mtDNA G4s), addressing limitations of such endpoint assays as DNA fragmentation index while requiring further standardization for G4 quantification protocols. The development of specialized fluorescent probes for these biomarkers enables multimodal assessment of sperm quality, forming the scientific foundation for this investigation. This work advances our preliminary dual-probe live-cell imaging platform,15 permitting concurrent spatial–temporal analysis of pHi and mtDNA G4s in motile spermatozoa. Contemporary advances in fluorophore engineering have established fluorescence microscopy as a cornerstone technology for real-time cellular signaling investigation.16 Current analytical tools (e.g., Fiji15 and CellProfiler17) predominantly employ manual region-of-interest (ROI) selection with intensity thresholding, while emerging deep learning approaches (e.g., c-ResUnet-based cell counting18 and StarDist-powered DL-SCAN1516) demonstrate enhanced capacity for automated signal quantification in crowded cellular environments. Nevertheless, existing methodologies exhibit critical limitations in sperm-specific analyses, particularly in correlating quantitative fluorescence measurements with distinct sperm compartments for diagnosis purpose.

This study introduces an innovative post-imaging analytical framework that integrates single-sperm fluorescence distribution mapping with functional biomarker quantification. Our system transcends conventional semen analysis by providing etiologically informative diagnostics through the synergistic application of molecular probes and computational imaging. This technological integration establishes a novel paradigm for male infertility diagnostics, offering unprecedented resolution in identifying pathophysiological mechanisms while maintaining clinical practicality. By bridging molecular diagnostics with assisted reproductive technology, this advance holds transformative potential for personalized etiology diagnosis and fertility management strategies.

PARTICIPANTS AND METHODS

Semen samples

The Ethics Committees of the West China Second University Hospital of Sichuan University (Chengdu, China) approved the study (Approval No. 111/2022). Each participant signed informed consent. The inclusion and exclusion criteria were according to the WHO guidelines (6^th edition, 2021).10 All semen samples were obtained by masturbation after 7 days of abstinence.

Fluorescent probe staining

Semen samples were washed three times with phosphate-buffered saline (PBS; Solarbio, Beijing, China) and then centrifuged at 300–600g for 10 min (5425R; Eppendorf, Hamburg, Germany). Then, G-IVF PLUS (Vitrolife, Gothenburg, Sweden) was used to obtain an appropriate concentration of sperm at 107 ml⁻¹ using the swim-up method. PATO was added to the sperm suspension at a concentration of 5 μmol l⁻¹ and AuNPs was at 75 mg ml⁻¹ and incubated at room temperature for 10 min.11,12 The spermatozoa were then washed three times with PBS and transferred to polylysine-coated NEST 3.5-mm glass-bottom dishes (P2100; Solarbio) for imaging with a laser confocal microscope (LSM780; Zeiss, Oberkochen, Germany).

Methodological framework

Deep learning-based detection and tracking techniques have achieved high accuracy on the public brightfield images or videos.19 However, there is currently no sufficient training data for our new staining technique. Hence, we propose an automatic analysis system, where a detection model trained on the public dataset is introduced to realize the fluorescent probe-stained semen segmentation, enabling the correlation of quantitative fluorescence measurements with precise subcellular localization patterns in distinct sperm compartments. In our framework, You Only Look Once version 8 (YOLOv8) was adopted for brightfield sperm detection, given its great success on the public dataset.19 Considering the diversity of fluorescent images from the use of different probes, automatic preprocessing and binary segmentation were applied. The detection result was then incorporated for instance segmentation, where stained spermatozoa were identified as individual segments for fluorescent quantification.

Brightfield image sperm detection

Our semen sample video incorporated both brightfield and fluorescent image sequences in pairs. Owing to the lack of labeled fluorescent data, spermatozoa were detected in brightfield images by using YOLOv8 trained on the public dataset. We focused on the sperm head region since it can be detected accurately owing to its distinct oval morphology (4–5 μm length) versus filamentous principal piece (45–50 μm) exhibiting rapid non-linear motility (>50 μm s⁻¹).

Our detection model was trained on the EVISEN dataset, which contains 6000 images from different donors and were taken at various magnifications for multisperm detection and tracking.20 For model training and testing, the dataset was split into a training set and a test set at a ratio of 9:1. The built-in mosaic and mix-up data augmentation methods of YOLOv8 were employed to enhance the robustness to real-world variations.21,22 The images were re-sized to the input size of 640 × 640 pixels.

Of the five models in the YOLOv8 series, YOLOv8s is well-balanced for speed and accuracy and is suitable for both central processing unit (CPU) and graphics processing unit (GPU) inference. YOLOv8s also achieved high small object detection accuracy by enhanced spatial pyramid pooling and an improved path aggregation network (PANet).23 Our sperm detection model was initialized with pre-trained weights of YOLOv8s on COCO201724 and then fine-tuned on our training dataset. This strategy incorporates prior knowledge from large-scale data, enhancing its training efficiency and performance.25

Model training and testing were conducted on a computer with an NVIDIA RTX 3090 GPU (24GB VRAM; NVIDIA Corporation, Santa Clara, CA, USA), an Intel Xeon Silver 4210R CPU (Intel Corporation, Santa Clara, CA, USA) and 128GB DDR4 RAM (Samsung Electronics Co., Ltd., Suwon-si, Korea), utilizing an Ubuntu 18.04 LTS operating system (Canonical Ltd., London, UK) with CUDA 11.0 acceleration (NVIDIA Corporation). The deep learning environment employed PyTorch 1.7.1 (Meta Platforms, Inc., Menlo Park, CA, USA) with cuDNN 8.0.5 (NVIDIA Corporation) and Python 3.7.12 (Python Software Foundation, Wilmington, DE, USA). The training parameters used were the official parameters of YOLOv8: 640 × 640 pixels input resolution, batch size 64 200 epochs (an epoch being when the entire training dataset passes through the model once), and stochastic gradient descent (SGD) optimization (momentum = 0.937, weight decay = 5 × 10⁻⁴) with cosine learning rate decay.26

Fluorescent image segmentation

Segmentation was applied to the corresponding fluorescent images to separate the marked sperm regions. Images were first preprocessed with a template to remove the scale bars. Subsequently, the fluorescent images were treated channel-wise where each channel represents one biomarker. In our experiment, the fluorescent images were of various intensities; we determined that images whose average intensity of all non-zero pixels below 3.5 were dark images, and a binarization operator with threshold equal to 3 (method 1) was applied to mark the foreground. For images with an average intensity over 3.5, the histogram equalization operation was first applied for image enhancement (method 2),27 and later, binarization was set with threshold 20. Note that histogram equalization was not applied to dark images to prevent noise amplification. Morphological methods were then used to separate the connected foreground images into individual regions, where each region represents the fluorescence marker of one individual spermatozoon or overlapping cluster.28

Subcellular fluorescence quantification

At the end, the segmentations of all channels were merged with the brightfield image to achieve subcellular fluorescence quantification. Each spermatozoon was located by a bounding box centered on the head; then, all the fluoresce segments of the same spermatozoon were identified by the overlapping areas with its bound box. Note that there could be multiple fluorescent regions intersecting with the sperm head bounding box, only the segment with the largest intersection area was selected. Our fluorescent probes marked both the sperm head and the whole cell; hence, the difference of the two segments was identified as the sperm principal piece. With the matching results, the fluorescent area and average fluorescence intensity of individual spermatozoon were calculated. Since detection can fail for the moving spermatozoa, multiple images captured from a microscope or multiple frames extracted from a video-recording were processed to obtain the final result.

Statistical analyses

All data are expressed as mean ± standard deviation (s.d.). GraphPad Prism software (version 8.0; GraphPad Software, San Diego, CA, USA) was used to analyze all statistics. Significant between- and among-group differences are indicated by P < 0.05.

RESULTS

Dual-probe fluorescence imaging system for mtDNA G4s and pHi co-mapping

The PATO-AuNP dual-probe system enabled concurrent visualization of mtDNA G4s and pHi through spectrally distinct fluorescence signatures (Figure 1a). PATO was synthesized via C8-position methylthiazole orange (mTO) conjugation and rotor modifications (Knoevenagel condensation; Figure 1b).12 Complementarily, the AuNP probe functionalized with rhodamine (acidic response: excitation wavelength [λex] 550 nm/emission wavelength [λem] 580 nm) and fluorescein (alkaline response: λex 450 nm/λem 510 nm)29 provided ratiometric pHi sensing (Figure 1c). Critical spectral separation was achieved with PATO emitting in the near infrared spectroscopy (NIR) window (λex 565 nm/λem >700 nm), showing <3% cross talk with AuNP channels (Figure 2a). Costained spermatozoa were imaged by confocal microscopy (405 nm/561 nm/640 nm lasers, GaAsP detectors) combined with differential interference contrast (DIC), followed by Fiji-based image fusion (Figure 2b). This integrated platform successfully resolved real-time mtDNA G4-pHi interplay during sperm motility, demonstrating its potential for molecular-level infertility diagnostics.

Figure 1.

(a) The schematic illustration of PATO-AuNP dual-probe system. (b) Conformational changes of PATO upon binding to mitochondrial DNA G-quadruplex (mtDNA G4s). (c) Structures of AuNPs at different values of pH. pHi: intracellular pH; Flu: fluorophore; Rh: rhodamine; Hy: hydroxyethyl; TA: thioctic acid; EDA: ethylenediamine.

Figure 2.

(a) The spectrum of AuNPs and PATO. (b) Fluorescence-labeled image of spermatozoa obtained via confocal fluorescence microscopy. mtDNA: mitochondrial DNA; Flu: fluorophore; Rh: rhodamine.

Integrated computational-imaging platform for sperm analysis

Building upon the dual-probe fluorescence co-mapping capability, we developed a computational platform integrating YOLOv8-based detection with fluorescent image segmentation (Figure 3a).

Figure 3.

(a) Schematic representation of probe staining and analysis based on YOLOv8. (b) The labeled image in EVISAN dataset. (c) The predict result of YOLOv8 on one of our brightfield images. (d) The comparison chart of two processing methods for fluorescent image. The upper image is dark (with an average grayscale value of 3.17), and the lower image is bright (with an average grayscale value of 9.13). The result of method 1 is on the left, and method 2 is on the right. (e) The combination of the results from sperm detection and fluorescence segmentation, after which stained spermatozoa are identified as individual segments for subcellular fluorescence quantification. The images show spermatozoa with red-green-yellow fluorescence from left to right: the upper show fluorescence segmentation and the lower show integrated detection-segmentation. (f) The flowchart of the system operation. YOLOv8: You Only Look Once version 8.

EVISAN dataset provides a suitable sperm pool with a degree of image heterogeneity and was annotated and checked by several biologists (Figure 3b). In our experiment, the comment index average precision (AP)30 was used to measure the detection accuracy, and the model trained on the EVISAN dataset achieved 95.4% AP on test set, showing precise and stable detection. The locations of detected sperm heads were labeled by bounding boxes (Figure 3c). The fluorescent images were processed by two pre-processing methods for different levels of brightness (Figure 3d). To achieve subcellular fluorescence quantification, a sperm-detection algorithm was initially applied to a brightfield image and the segmentation applied to the corresponding single-channel fluorescent images. The detection and segmentation results were combined for subcellular fluorescence quantification (Figure 3e), where overlapping bounding boxes and fluorescent segments were paired to identify the same spermatozoon. The fluorescent area and average fluorescent intensity of sperm head and principal piece were then calculated (Table 1). The current algorithm divides spermatozoa into two regions: the head (including the midpiece) and the principal piece (main segment). Owing to technical limitations, the midpiece mitochondrial sheath was temporarily incorporated into the head for fluorescence quantification, and the midpiece specific segmentation will be improved in the future.

Table 1.

Colocalization and quantitative statistical analysis of fluorescence image recognition data

Color	Sperm ID	Part	Area (pixel)	Average fluorescence intensity
Green	1	Head	314	7.05
		Principal piece	43	7.51
	2	Head	416	10.64
		Principal piece	137	12.74
Red	1	Head	127	2.9
		Principal piece	0	0
	2	Head	204	9.71
		Principal piece	9	6.11
Yellow	1	Head	140	10.84
		Principal piece	0	0
	2	Head	188	26.38
		Principal piece	10	14.3

Our analysis system demo is available at the link: http://106.52.98.34/, and source code for sperm detection is available at the link: https://github.com/1377Zing/Yolov8_Sperm_detection.git (last accessed on 2025 April 30). The source code for our demonstration, which also includes instructions on system usage, can be found at: https://github.com/1377Zing/yolo-segment_flask.git (last accessed on 2025 April 30). The source code for our sperm detection algorithm is available at: https://github.com/1377Zing/Yolov8_Sperm_detection.git (last accessed on 2025 April 30). The system enables users to upload images and automatically generate detailed analysis results (Figure 3f).

Subcellular fluorescence profiling and diagnostic potential

Spermatozoa can be stained by fluorescent probes with different fluorescence of intracellular areas, which contributes to detecting their individuality and specificity (Figure 4a). The automatic localization on individual spermatozoon of fluorescent spots enables quantitative analysis of the heads and principal pieces (identifying fluorescent spots as principal piece positions after excluding the head positions; Figure 4b), including the fluorescence area and the average fluorescence intensity (Table 1).

Figure 4.

(a) After probe staining, each spermatozoa shows different fluorescence results. (b) Identify and delineate the head and principal piece regions of spermatozoa using confocal fluorescence imaging, followed by a quantitative analysis of the resulting fluorescence data. (c) Quantitative assessment and comparative analysis of average fluorescence intensity in sperm head and principal piece regions. (d) Fluorescence intensity ratio of the head/tail (principal piece) of each sperm for each clinical case. Cases 1-4 were collected from clinics and their semen parameters were normal according to clinic diagnosis.

When the software was established, we collected another six clinical cases and detected sperm pHi and mtDNA G4s with our fluorescence probes and confocal microscopy, quantifying fluorescence to assess intensity variations within spermatozoa (Figure 4c) that display different marker concentrations in head and principal piece regions (Figure 4d). The analysis data for the multi-color co-staining of semen samples is provided, which encompasses six clinical cases: https://github.com/1377Zing/results_of_sperm.git (last accessed on 2025 April 30).

DISCUSSION

The integration of the detection algorithm and fluorescence segmentation enables subcellular fluorescence quantification in live spermatozoa. It solves the persistent challenges in live sperm analysis, especially in the segmentation of sperm principal pieces, in a concise and low-computational-intensity manner. This computational framework, when integrated with confocal–brightfield multimodal imaging, enables precise subcellular fluorescence mapping, a critical advance beyond flow cytometry’s limitations in spatial resolution. While flow cytometry permits population-level pHi/G4 quantification, it fails to capture compartment-specific dynamics essential for understanding capacitation mechanisms, as indicated by our findings of pHi polarization and mtDNA G4 mitochondrial clustering. These spatially resolved profiles are correlated strongly with functional outcome, underscoring their diagnostic potential.

Our study introduces an AI-enhanced computational-imaging platform for multimodal sperm analysis, integrating dual-probe fluorescence co-mapping with spatially resolved functional biomarker quantification. A key innovation lies in the methodological framework that reconciles live-cell fluorescence imaging with computational precision. By anchoring spermatozoa to polylysine-coated surfaces, we achieved stable imaging conditions essential for high-resolution confocal microscopy while preserving flagellar motility. Although this approach restricts progressive sperm movement, rendering conventional CASA kinematic parameters unobtainable, it enables robust subcellular fluorescence profiling of dynamic biomarkers (pHi and mtDNA G4s) in motile spermatozoa. This trade-off underscores the platform’s unique value in bridging molecular diagnostics with structural–metabolic evaluation, transcending the limitations of endpoint assays like DNA fragmentation index. This holistic view demonstrates unique structural and molecular characteristics could improve the understanding of sperm structure–function relationships and have profound implications for reproductive biology, affecting fertility assessments, diagnostics, and the advances in reproductive technologies. This suggests the potential application of the algorithm in clinical detection, particularly in live sperm fluorescence imaging.

This study presents an integrative approach combining YOLOv8-based detection with adaptive fluorescence segmentation, offering a novel solution for sperm-specific image analysis. The framework effectively addresses the long-standing challenge of signal resolution in high-density sperm populations, particularly in clinical samples with heterogeneous motility patterns. While this methodology significantly enhances spatial resolution for overlapping fluorescent signals, its limitations must be acknowledged: (1) reliance on static imaging restricts real-time capture of dynamic biomarker fluctuations during sperm motility and (2) current architecture cannot resolve the midpiece–principal piece junction in hyperactivated spermatozoa, where rapid flagellar undulation exceeds temporal resolution thresholds. Future research could incorporate transformer-based architectures for real-time flagellar kinematic profiling and integrate multiethnic training cohorts to improve cross-population generalizability.

This platform establishes the scalable framework for molecular phenotyping of sperm functional competence, unifying three critical dimensions: spatially resolved biomarker quantification, AI-driven subcellular localization mapping, and mechanistic correlation with fertilization potential. By synchronizing live-cell fluorescence dynamics with deep learning analytics, it transcends conventional motility-based binary assessments (e.g., viability thresholds) to provide etiological stratification for idiopathic male infertility, a pivotal advance toward precision reproductive medicine.

However, methodological challenges remain. First, while the EVISAN dataset represents a pioneering effort, its resolution and staining diversity are insufficient for robust quantification of G4/pHi biomarkers. A multi-institutional consortium should curate high-quality datasets using standardized probe protocols. Second, the current head-centric analysis simplifies fluorescence quantification but overlooks energy metabolism in flagellar motion. Emerging microparticle tracking algorithms could bridge this gap by correlating midpiece metabolic outputs (e.g., mtDNA G4 dynamics) with flagellar beat patterns. Finally, existing biomarker thresholds remain cohort-specific, necessitating a WHO-coordinated multicenter consortium to establish clinically validated reference ranges across etiologies (e.g., oxidative stress, genetic defects, and epigenetic anomalies).

Notably, this platform creates a diagnostic-intervention closed-loop ecosystem through the integration of molecular andrology and assisted reproductive technologies. AI-derived biomarker profiles may guide personalized sperm selection for intracytoplasmic sperm injection (ICSI), while post-IVF outcomes can iteratively refine algorithmic training. This bidirectional framework positions the platform not merely as an analytical tool but as a catalyst for transforming infertility management, from reactive treatment to mechanism-driven prevention. Future work should prioritize large-scale clinical validation, algorithmic optimization, and exploration of multidimensional applications in translational reproductive medicine.

In summary, our study not only offers a novel methodology for sperm quality assessment but also lays the groundwork for future research in biomedical image analysis. By continuously optimizing our model and expanding the dataset with our samples, we anticipate making significant advancements with the novel diagnosis tool in reproductive medicine.

AUTHOR CONTRIBUTIONS

WMX and AB conceived this project. YZW and SXW carried out the experiments. XDS and WWZ collected clinical information of the patients. KKY designed and synthesized fluorescent probes. YXN and HYZ developed the algorithm and software. YZW, YXN, and YXC drafted the manuscript. WMX, AB, and XXY reviewed and edited the manuscript. All authors read and approved the final manuscript.

COMPETING INTERESTS

All authors declare no competing interests.

SUPPLEMENTARY MATERIALS AND METHODS

The source code for our demonstration, which also includes instructions on system usage, can be found at: https://github.com/1377Zing/yolo-segment_flask.git.

The source code for our sperm detection algorithm is available at:

https://github.com/1377Zing/Yolov8_Sperm_detection.git.

We also provide the analysis data for the multi - color co - staining of semen samples, which encompasses six clinical cases: https://github.com/1377Zing/results_of_sperm.git

ETHICAL APPROVAL

All procedures performed were in accordance and with the Sixth Edition of the WHO Laboratory Manual for the Examination and Processing of Human Semen and human semen dataset was obtained from Zenodo that is a general-purpose open repository developed under the European OpenAIRE program and operated by CERN.

INFORMED CONSENT

Informed consent is not available due to human semen dataset was obtained from Zenodo that is a general-purpose open repository developed under the European OpenAIRE program and operated by CERN.