Deep learning-based approach for sperm morphology analysis

Abstract

Male infertility is a highly prevalent condition throughout the world. Sperm morphology analysis（SMA） is one of most important examination for evaluating male infertility. This paper highlights the strengths, limitations, and clinical applicability of conventional machine learning (ML) models and deep learning (DL) models in SMA from various studies. Simultaneously, we explore the potential role of segmentation and classification of complete sperm structure based on deep learning algorithms. Therefore, this narrative literature review aims to summarize the current evidence of artificial intelligence and machine learning applications for sperm morphology analysis and explore further recommendations about deep learning algorithms applications to practically enhance the performance.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12894-025-01946-w.

Background

Infertility has become an urgent global public health issue, and male factors contribute to about 50% of infertility cases [1]. The quality of male semen shows a declining trend over time in sperm concentration and sperm numeration [2]. Research has shown that this trend is especially pronounced among young men in China, notably in terms of sperm morphology, vitality, and quantity [3, 4]. Clinically, most sperm quality issues are caused by abnormal sperm morphology [5]. Therefore, sperm morphology analysis is a crucial laboratory test in male fertility assessment. Clinicians comprehensively evaluate sperm quality by analyzing proportion of abnormal morphology sperm in a fixed number of over 200 sperms and the specific types of sperm defects [6]. These results not only predict natural pregnancy outcomes but also provide more diagnostic information for male testicular and epididymal function [7].

Sperm morphology analysis represents a significant challenge in morphological analysis, characterized by high recognition difficulty [8]. According to the classification standards established by the World Health Organization (WHO), sperm morphology is divided into the head, neck, and tail, with 26 types of abnormal morphology, requiring the analysis and counting of more than 200 sperms [9]. An issue worth noting is that manual observation involves a substantial workload and always influenced by the subjectivity of observers, thereby hindering clinical diagnosis of male infertility by physicians [10]. Thus, morphological evaluation of sperm still faces considerable limitations in reproducibility and objectivity. Currently, applying machine learning (ML) algorithms to automatic recognition of clinical morphology, including urine and blood, have demonstrated a notably success. However, conventional algorithms heavily relied on manual feature extraction of human sperm cells. To overcome these limitations of conventional machine learning models, recent studies have shifted toward deep learning algorithms. Consequently, to address the various challenges associated with SMA, it is essential to build a sperm automatic recognition system based on DL algorithms. Critically, this system should demonstrate two key points: (1) accurate automated segmentation of sperm morphological structures (head, neck, and tail), and (2) substantial improvements in the efficiency and accuracy of sperm morphology analysis [11, 12].

This article presents a comprehensive review which integrates the current state of ML technology in sperm morphology analysis both domestically and internationally, providing more detailed literature support for research on automated sperm recognition systems. In detail, it encompasses a detailed analysis of the lack of standardized, high-quality annotated datasets, limitations of conventional machine learning algorithms and development potential of deep learning algorithms. Over the long term, the review lays the foundation for improving the efficiency of sperm quality analysis and enhancing the diagnostic efficiency of infertility-related diseases.

The aim of the review is to show research progress on the application of artificial intelligence techniques in sperm quality assessment, with a specific focus in sperm morphology. A comprehensive search was conducted in PubMed Central, Web of Scienceand, MEDLINE-Academic databases to identify relevant articles. A literature research was performed using the terms “Human Infertility”, “sperm morphology”, “sperm AI” and “sperm dataset” as key words between 2010 and 2025. We limited our search to peer-reviewed articles published in English.

Lack of standardized, high-quality annotated datasets

In recent years, with the rapid development of computer technology, utilizing high-performance computers to process large volumes of medical images has become a significant focus in medical big data research. Recent studies have shown that, compared to conventional machine learning techniques, deep learning relies on big data for multidimensional data extraction and analysis, enabling the automatic extraction of features and training [13]. Deep learning has shown promising prospects in the field of automatic recognition of sperm morphology. To ensure the effective application of deep learning in sperm morphology research, attention must still be paid to the quality and diversity of datasets to guarantee the generalization ability of the model [14]. Table 1. Related research has trained using the free and public datasets to evaluate sperm images [15–23], such as HSMA-DS: Human Sperm Morphology Analysis DataSet [15], MHSMA: Modified Human Sperm Morphology Analysis Dataset [18] and VISEM-Tracking [23]. Among these, Javadi S et al. [18] proposed a deep learning model extracted features such as acrosome, head shape, and vacuoles from a dataset named MHSMA, of 1,540 images of different types of sperm. Nevertheless, limitations in the dataset including low resolution, limited sample size and insufficient categories are still exist. In more newly study, Chen A et al. [22] established the SVIA (Sperm Videos and Images Analysis) dataset, comprising: (i) 125,000 annotated instances for object detection, (ii) 26,000 segmentation masks, and (iii) 125,880 cropped image objects for classification tasks. Consequently, the inherent complexity of sperm morphology, particularly the structural variations in head, neck, and tail compartments, presents fundamental challenges for developing robust automated analysis systems across existing datasets. From previous experience, there is evidence that building standardized, high-quality annotated dataset is a challenging task [18, 23, 24]. First of all, many medical institutions still primarily rely on conventional sperm assessment methods. On the one hand, the visual analysis of sperm morphology often involve subjectivity and reproducibility. On the other hand, conventional methods result in many valuable image data being unable to be systematically saved, leading to data loss [22, 24]. Moreover, sperm may appear intertwined in the images, or only partial structures may be displayed due to being at the edges of the image, which affects the accuracy of image acquisition and also increases the difficulty of subsequent data analysis. Notably, sperm defect assessment under microscopy requires simultaneous evaluation of head, vacuoles, midpiece, and tail abnormalities, which substantially increases annotation difficulty [25, 26]. Consequently, to make deep learning algorithms better use in sperm automatic recognition systems, we should focus on establishing a series of standardize processes for sperm morphology slide preparation, staining, image acquisition, and annotation. In other words, the larger and more high-quality data become available, the better the performance of automatic recognition systems will be.

Table 1.

Free and public datasets to evaluate sperm images

Study	Time	Name		Ground truth	Images
Ghasemian F et al.[15]	2015	HSMA-DS: Human Sperm Morphology Analysis DataSet	Non-stained, noisy, and low resolution	Classification	1457 sperm images from 235 patients, unstained sperms.
Chang V et al.[16]	2017	SCIAN-MorphoSpermGS	Stained and possess higher resolution	Classification	1854 sperm images,the data was classified into five classes: normal, tapered, pyriform, small, and amorphous.
Shaker F et al.[17]	2017	HuSHeM: Human Sperm Head Morphology	Stained and possess higher resolution	Classification	725 images, but only 216 images containing sperm heads are publicly available.
Javadi S et al.[18]	2019	MHSMA: Modified Human Sperm Morphology Analysis Dataset	Non-stained, noisy, and low resolution	Classification	1540 grayscale sperm heads images.
McCallum C et al.[19]	2019	—	Bright-field sperm	Classification	1064 cropped images of six healthy participants.
Saadat H et al.[20]	2019	VISEM: a multimodal video dataset of human spermatozoa	Low-resolution unstained grayscale sperm and videos	Regression	A multi-modal dataset containing different data sources such as videos, biological analysis data, and participant data from 85 different participants.
Ilhan HO et al.[21]	2020	SMIDS: Sperm Morphology Image Data Set	Stained sperm images	Classification	3000 images with three classes.1005 abnormal, 974 non-sperm, and 1021normal sperm head images.
Chen A et al.[22]	2022	SVIA: Sperm Videos and Images Analysis dataset	Low-resolution unstained grayscale sperm and videos	Detection, segmentation and classification	4,041 low-resolution images of unstained sperm and videos.
Thambawita V,et al.[23]	2023	VISEM-Tracking	Low-resolution unstained grayscale sperm and videos	Detection, tracking, and regression	656,334 annotated objects with tracking details.

Conventional ML algorithms with limited performance

Machine learning (ML), a specialized branch of artificial intelligence (AI) with demonstrated capabilities in large-scale data processing, has Extensively applied in other medical imaging fields. Similarily, ML substantially alleviating analytical workload while significantly reducing inter-observer variability in sperm quality assessment. A summary of the conventional ML algorithms assessing sperm morphology is presented in Table 2 [16, 27–31]. These automatic techniques for analyzing human sperm morphology are essential to avoid human errors and variation in results. Among conventional machine learning approaches, K-means, support vector machine (SVM) and decision trees are the most archetypal algorithms. But they are fundamentally limited by their non-hierarchical structures and handcrafted features. Specifically, these methods highly rely on manually designed image features (e.g., grayscale intensity, edge detection, and contour analysis) for effective sperm image segmentation.

Table 2.

Summary of conventional ML in SMA models

In the field of sperm morphology classification, conventional ML algorithms demonstrated considerable success. This achievement is realized through a standardized ML pipeline designed to differentiate normal and abnormal morphological features from individual sperm images. Firstly, shape-based descriptors and other feature engineering techniques were used for manual extraction of sperm cell features. Then, a classifier was used to classify a sperm image, such as a SVM or a neural network. Bijar A et al. [27] proposed a Bayesian Density Estimation-based model achieving 90% accuracy in classifying sperm heads into four morphological categories (normal, tapered, pyriform, and small/amorphous). However, the current model’s detection of normal sperm relies exclusively on shape-based morphological labeling and classification. Expanding feature extraction to include texture, depth, and grayscale data should be added to further research. For the segmentation of stained sperm images, Chang V et al. [29] proposed a two-stage framework that locates the sperm head using the k-means clustering algorithm, and combines clustering with histogram statistical methods for segmentation. Additionally, they also explored various combinations of color spaces, further enhancing the accuracy of segmentation for the sperm acrosome and nucleus [29]. There are other detection and segmentation algorithms such as Hu moments, Zernike moments, and Fourier descriptors, combined with K-neighbor, simple Bayes and decision tree [27, 29, 31]. Works done by Tseng KK et al. [28], Chang V et al. [16], Mirsky SK et al. [30] and ZONYFAR C et al. [31] represented more precise classification of sperm head. Support vector machines is the most commonly algorithm in the conventional ML. Mirsky SK et al. [30] trained a support vector machine (SVM) classifier to to classify sperm heads good and bad based on Over 1400 human sperm cells from 8 donors. The SVM model yielded strong discriminatory power. Evaluation metrics confirmed the model’s diagnostic efficacy: 88.59% area under the receiver operating characteristic curve (AUC-ROC), 88.67% area under the precision-recall curve (AUC-PR), and precision rates consistently above 90%. ZONYFAR et al. [31] developed a sperm head detection model based on Bayesian density estimation, Hu moments, Zernike moments [32] and Fourier descriptors, achieving a high classification accuracy of 90%. In contrast, Chang V et al. [16] highlighted the limitations of conventional methods, confirming the high degree of inter-expert variability in the SMA. Their approach utilizing the Fourier descriptor and SVM among non-normal sperm heads achieved classification accuracy of only 49%.

Although these conventional algorithms have significantly promoted the development of sperm morphology analysis and laid the foundation for subsequent research. But most of them only discussed the accuracy of sperm head classification, without covering various categories of head, neck and tail. Moreover, most studies only classify the sperm head as normal or abnormal, without being able to detect the complete structural morphology of sperm [33, 34]. Additionally, there are still technical difficulties in the analysis of sperm morphology to correctly distinguish the sperm head from the impurities in semen fragments. They only rely on image features such as thresholds and textures, which can be somewhat inadequate in terms of both segmentation accuracy and efficiency, often resulting in over-segmentation or under-segmentation issues [34]. Therefore, in order to further improve segmentation accuracy, many researches turned to rely on manual extraction of image features. However, this method of feature extraction is not only cumbersome and time-consuming but often reduces the generalization ability of the algorithms, leading to significantly different performance of the same algorithm across different datasets [35].

In summary, more available sperm imaging datasets and better machine learning algorithms are the key in this process, similarly to other automated recognition systems development [30]. There is an urgent need to explore more advanced algorithms to improve the accuracy and efficiency of analysis, providing more reliable tools for the study of sperm morphology.

DL algorithms with new breakthrough

Deep learning (DL), as an advanced form of machine learning paradigm, has demonstrated significant advantages over conventional approaches in sperm morphology analysis. Three critical advancements distinguish this approach: first, the ability to process raw images without extensive preprocessing or manual feature engineering; second, the potential for achieving higher accuracy from a large of similar medical images studies; third, the visualization with unprecedented visibility into discriminative sperm morphology features [36, 37]. Table 3 [18, 21, 36–40] summarizes recent DL applications in sperm morphology analysis, which have showed significant improvement in segmentation and classification over other machine learning approaches [41, 42]. The HuSHeM and SCIAN datasets represent the most widely adopted, particularly for evaluating deep learning-based sperm classification structure [13, 17–20]. The most commonly used deep learning algorithms currently include convolutional neural network(CNN), deep neural network(DNN), deep-convolutional neural network (D-CNN).

Table 3.

Summary of DL in SMA models

Segmentation

Segmentation is an essential part of sperm morphology analysis. While prior studies have attempted automated sperm head segmentation, current ML approaches still face three major limitations [21, 43, 44]. On the one hand, imaging artifacts such as microscope optical reflections, sperm occlusions, and halos affect the detection accuracy. On the other hand, the existence of non-sp45erm particles with morphological similarity to normal sperm was still a problem. Substantial studies only focused on the size information of extracted parts, thereby failing to effectively distinguish sperm from similarly-shaped debris. Furthermore, precise acrosome-nucleus separation presents particular difficulties. The acrosome segmentation challenge is particularly pronounced due to intensity similarity between acrosome and adjacent head regions under varying illumination, ultimately degrading segmentation performance [45].

To increase the segmentation performance, there are several studies have make a great deal of success. Shaker F et al. [45] proposed an edge-based active contour method and a novel tail point detection method, which refined the segmentation by locating and removing the midpiece from the segmented head. The method extracted the outer contour of the head more precisely than the previous methods. But there showed lower performance in segmentation of acrosome. The reason is that the segmentation of acrosome is done based on the difference in the intensity. Perdrix et al. [46] proposed a deep learning framework for automated human sperm head segmentation. Their approach combined a deep convolutional neural network (DCNN) for head segmentation with a support vector machine (SVM) for classifying each pixel of segmented heads to nucleus and acrosome regions. As a result, Dice Similarity Coefficient for the head, the acrosome, and the nucleus segments were obtained 0.94, 0.87, and 0.88 respectively. However, the proposed method is validated on the expert delineated dataset with only 20 images of human semen smears.

Currently, research on the application of deep learning in sperm morphology analysis mostly focuses on semantic segmentation of sperm head. Melendez R et al. [47] used the semantic segmentation network U Net for sperm semantic segmentation, achieving an 88% IOU score and 94% DICE score on the annotated sperm segmentation dataset. Ilhan et al. [21] developed an automated sperm segmentation framework, including a group-sparse signal named as modified overlapping group shrinkage (MOGS) to denoise, combining regions of interest (ROI) segmentation techniques and the Fuzzy C-Means clustering approach. This integrated approach enabled precise morphological segmentation of individual sperm within dense populations. The application of clustering techniques after MOGS preprocessing yielded superior region-specific segmentation, enabling more precise segmentation of sperm/non-sperm regions. While the proposed system demonstrates potential utility in clinical laboratory settings, further refinement of the feature extraction remains necessary. Additionally, segmentation algorithms based on convolutional neural networks often use inherent downsampling mechanisms to extract contextual information from images, which frequently leads to the loss of detailed information in the images. Particularly small objects such as sperm heads are especially prone to the problem of detail loss.

Furthermore, in the complete morphological analysis of sperm, defects in the middle and main segments of sperm cannot be ignored. While pursuing high precision, it is essential to consider the lightweight processing of the network model to achieve higher operational efficiency and a more reasonable resource utilization rate. In summary, for the task of segmenting the sperm head, we must focus on preserving detailed information and strive for a balance in network design to enhance both efficiency and performance.

Classification

While conventional machine learning algorithms have demonstrated considerable efficacy in sperm classification tasks, their dependence on manually engineered feature extraction remains a fundamental limitation [48]. This human-engineered approach, characterized by time-consuming and subjectivity. Deep learning (DL) has emerged as a viable solution to these limitations by automating feature learning directly from raw images. For image classification tasks, convolutional neural networks (CNNs) are the most clinically validated and routinely employed of deep learning algorithms. Despite recent advances, accurate classification of sperm morphological defects remains three critical limitations. Firstly, there are still lack of high-quality annotated datasets. Secondly, DL algorithms are heavily depend on various datasets for training, such as HuSHeM [24], SCIAN [16], and MHSMA [17]. However, training on different datasets often results in significant differences in accuracy. Last but most importantly, existing research on sperm morphology based on deep learning algorithms generally suffers from insufficient practical effectiveness, primarily reflected in its inadequate ability to recognize different parts of the sperm.

In the task of classifying sperm morphology defects, in a more recent study, Iqbal et al. [49] developed a customized CNN architecture as an alternative to conventional pre-trained networks, such as VGG16. They comfired that using fewer filters and parameters in the CNN architecture achieved classification accuracies of 63% on SCIAN-Morpho and 95% on HuSHeM datasets. Notably, the implementation required manual intervention during cross-validation folding. These findings underscore that there were significant differences in classification performance across different datasets. Following this study, Abbas A et al. [38] introduced the concept of multi-task learning to the field of SMA in the first time. They combined Deep Multi-task Transfer Learning (DMTL) with the classification of sperm head defects, making full use of the highly correlated characteristics among the sperm acrosome, head, and vacuoles. They reached the accuracy of 84.00%, 80.66%, and 94.00% on the head, acrosome, and vacuole, respectively, achieving better results in comparison with other study [50]. Notably, to better distinguish between sperm and impurities, Mahali MI et al. [51] proposed a deep learning fusion architecture, called SwinMobile, which combines the shifted windows vision transformer (Swin) and MobileNetV3 to classify sperm and impurities in SVIA Subset-C. To suppress automated noise, they further enhanced this framework by incorporating an autoencoder (SwinMobile-AE). In addition, experimental results demonstrated that SwinMobile-AE has the most strong classification performance across multiple datasets (SVIA, HuSHeM, and SMIDS) compared other models.

Despite more accurate results that previous studies have gained, additional work remains to be done to reach higher values in the deep learning models of sperm classification. Firstly, there are still a various of problems which are not yet completely solved both in algorithms and models building, and further research needs to be conducted to improve these methods. Secondly, to obtain better generalizability, it is strongly recommended to expand the image dataset in the future, including sample collection through multi-center collaborations, combining unlabeled data with partially labeled data.

Conclusions

In summary, the morphological analysis of sperm plays an indispensable role in the male sperm quality evaluating. This article summarizes the current research status from sperm segmentation to sperm morphological defect classification, based on existing research in the field of sperm imaging. At present, in terms of the practical difficulties faced in the field of sperm morphology analysis, the following issues still require further research in the future. First, researchers and clinicians into building a large capacity, free and open standard sperm morphology image dataset. Second, the current method only focus on segmenting the head of sperm. Future development is required to establish reliable segmentation techniques for the midpiece and tail regions, which are critical for comprehensive sperm defect identification. Third, future research should explore novel deep learning techniques to enhance classification performance across diverse datasets, addressing current limitations in generalizability. Ultimately, it is essential to overcome technical barriers posed by deep learning adoption and guarantee the safe and accountable application of deep learning in clinical sperm analysis.

Abbreviations

AI

artifcial intelligence

SMA

sperm morphology analysis

ML

machine learning

DL

deep learning

D-CNN

deep-convolutional neural network

DNN

Deep neural network

HSMA-DS

Human Sperm Morphology Analysis DataSet

HuSHeM

Human Sperm Head Morphology

MHSMA

Modified Human Sperm Morphology Analysis Dataset

VISEM

a multimodal video dataset of human spermatozoa

SMIDS

Sperm Morphology Image Data Set

SVIA

Sperm Videos and Images Analysis dataset

Authors’ contributions

Authors’ contributions：(I) Conception and design: Bianping Liang, Mingxue Wang; (II) Collection and assembly of references: Bianping Liang, Mingxue Wang; (III) Analysis and interpretation of references: Bianping Liang, Mingxue Wang; (IV)Manuscript writing: Bianping Liang, Mingxue Wang; (V)Final approval of manuscript: Bianping Liang, Mingxue Wang. Bianping Liang and Mingxue Wang have contributed equally to this work. Corresponding author: Bianping Liang.

Funding

Not applicable.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.