Deep-learning based model for sperm morphology assessment using the SMD/MSS dataset

Abstract

Background

Manual sperm morphology assessment is recognized as a challenging parameter to standardize due to its subjective nature, often reliant on the operator’s expertise. Our study aims to address this issue by developing a predictive model for sperm morphological evaluation utilizing artificial neural networks trained on the SMD/MSS(Sperm Morphology Dataset/Medical School of Sfax) dataset enhanced through data augmentation techniques.

Methods

A total of 1000 images of individual spermatozoa were acquired using the MMC CASA system. Expert classification, based on the modified David classification for sperm morphology, was conducted by three experts. Data augmentation techniques were employed to augment the database. Subsequently, an algorithm utilizing a Convolutional Neural Network (CNN) was created, trained, and tested for spermatozoa classification.

Results

SMD/MSS dataset, initially comprised 1000 images and extended to 6035 images after the application of data augmentation techniques. The deep learning model produced satisfactory results, with an accuracy ranging from 55% to 92%.

Conclusions

Our deep learning approach for sperm morphology classification enables the automation, standardization, and acceleration of semen analysis. It underscores the significance of artificial intelligence in medical applications, with a particular focus on its impact in the field of reproductive biology.

PLAIN LANGUAGE SUMMARY

Infertility affects many couples, and one important test used to study male fertility is the examination of sperm under a microscope. Traditionally, experts look at the shape of sperm to decide whether they are normal or abnormal. This process is important but can be slow and sometimes differs between laboratories. In this study, we created a new image database of sperm. The images were carefully classified by three experts, including sperm with different head, midpiece, and tail problems. To make the database more balanced, we used computer methods to create additional images. We then trained a computer program, based on artificial intelligence, to recognize the different sperm shapes. The results showed that the program could reach a level of accuracy that is close to expert judgment. Our findings suggest that artificial intelligence may help make sperm analysis more reliable and consistent in the future, which could benefit couples undergoing fertility testing.

ARTICLE HIGHLIGHTS

• The development of SMD/MSS (Sperm Morphology Dataset/Medical School of Sfax) dataset.

• The dataset includes normal and abnormal spermatozoa, covering head, midpiece, and tail anomalies according to the modified David classification.

• Images were augmented using multiple techniques to balance morphological classes.

• A deep learning algorithm (Python 3.8) was trained and evaluated on the dataset.

• The algorithm achieved promising accuracy compared with expert classifications.

• The study highlights the potential of AI to improve sperm morphology assessment in clinical practice

1. Introduction

Infertility is defined as the absence of conception after 12 months of regular and unprotected sexual intercourse [1,2]. It has become a significant public health issue [3], affecting nearly 15% of couples [4]. In approximately half of cases, male factors are involved, either alone (20 to 30%) or in combination with female factors [5].

Semen analysis is the key test during the initial investigation of male infertility. Among the assessed semen parameters, sperm morphology requires special attention as it is considered of greatest clinical interest and most correlated with fertility [6–8]. Sperm morphology is assessed in the majority of laboratories using the manual method [9], as outlined in the WHO (World Health Organization) manual for semen analysis [10]. Several recommendations and technical procedures are described in order to standardize the practice of this examination. Furthermore, internal (IQC) and external (EQC) quality controls are highly recommended to improve performance.

Despite these efforts, the manual assessment of sperm morphology remains highly subjective, challenging to teach, and strongly dependent on the technician’s experience [7]. In an attempt to address these issues, the Computer-Assisted Semen Analysis (CASA) method was developed. This system allows the sequential acquisition of images using a microscope equipped with a camera [11]. However, its routine use in automated sperm morphology analysis remains limited for several reasons [12]. It has a limited ability to accurately distinguish between spermatozoa and cellular debris and to classify midpiece and tail abnormalities. Also, the limited quality of the captured microscopic images often leads to unsatisfactory results [13,14].

Artificial intelligence (AI)-based image-processing techniques have been proposed to enhance performance. However, the robustness of these technologies relies primarily on the creation of a large and diverse database. Considerable progress has been made in developing databases with WHO classification [15] or Kruger classification (strict WHO 2010 criteria) [3]. Yet, to our knowledge, there is still no published dataset on automated morphological assessment according to David’s classification, which is used by a large number of laboratories worldwide. Furthermore, two major issues were frequently encountered during the construction of databases: the limited number of images, and the heterogeneous representation of the different morphological classes. Hence, database augmentation techniques were suggested to compensate for these shortcomings [16,17].

The objectives of our study were:

• to develop SMD/MSS (Sperm Morphology Dataset/Medical School of Sfax) dataset and to and enhance its power using data augmentation techniques

• to develop a predictive model for sperm morphological classification based on artificial neural networks.

2. Methods

2.1. The proposed dataset

2.1.1. Sample preparation

This prospective study was conducted at the Laboratory of Reproductive Biology, Medical School of Sfax, Tunisia. Smears were prepared from semen samples obtained from 37 patients attending the laboratory for semen analysis after obtaining informed consent. We included samples with a sperm concentration of at least 5 million/mL and varying morphological profiles to maximize examples of different morphological classes. Samples with high concentrations (>200 million/mL) were excluded to avoid image overlap and to facilitate the capture of whole sperm. On average, 37 ± 5 images were captured per sample, depending on the density of the sample and the distribution of spermatozoa on the smear. The smears were prepared following the guidelines outlined in the WHO manual [1] and were stained with RAL Diagnostics staining kit.

2.1.2. Data acquisition step

The MMC CASA system was employed in our study for acquiring images from sperm smears. It consists of an optical microscope equipped with a digital camera, facilitating the acquisition and the storage of images from sperm smears. We used bright field mode with an oil immersion x100 objective. The CASA morphometric tool accurately determined the width and length of the head, as well as the length of the tail for each spermatozoon. Each image in our dataset contains a single spermatozoon, comprising a head, a midpiece, and a tail.

2.1.3. Classification

Each spermatozoon underwent manual classification by three experts from our laboratory, each possessing extensive experience in semen analysis. The classification was conducted following the modified David classification [18], which includes 12 classes of morphological defects:

• 7 head defects: tapered (a), thin (b), microcephalous (c), macrocephalous (d), multiple (e), abnormal post-acrosomal region (f), abnormal acrosome (g),

• 2 midpiece defects: cytoplasmic droplet (h), bent (j),

• 3 tail defects: coiled (n), short (l), multiple (o).

Each spermatozoon was independently classified by the three experts. An Excel spreadsheet was created and shared among experts, each of whom having a dedicated section to document the various morphological classes for each part of the spermatozoon.

2.1.4. Image labeling

For each image, an image file name was assigned containing an uppercase letter indicating the type of anomaly along with a number representing the spermatozoon’s identification for each class:

• A: Tapered,

• B: Thin,

• C: Microcephalous,

• D: Macrocephalous,

• F: Abnormal post-acrosomal region

• G: Abnormal acrosome,

• E: Multiple heads,

• H: Cytoplasmic droplet,

• J: Bent,

• N: Coiled tail,

• L: Short tail,

• O: Multiple tails,

• CN: Associated anomalies,

• NR: Normal,

A ground truth file has been compiled for each image in the dataset. This file includes the image name, folder number, classification by the three experts and the dimensions of the sperm head and tail (Figure 1). If the spermatozoon contains associated anomalies (CN), the various sperm anomalies are detailed in this ground truth file.

Figure 1.

Image labeling in the proposed dataset. (a) Sperm image. (b) Corresponding ground truth file.

2.2. Inter expert agreement

A very important aspect of our study is the analysis of the inter-expert agreement distribution. The underlying complexity of the sperm cell classification task can be studied by evaluating the degree of agreement between different experts. There were three separate agreement scenarios among the three experts: 1: No agreement (NA) among the experts; 2: partial agreement (PA): 2/3 experts agree on the same label for at least one category, and 3: total agreement (TA): 3/3 experts agree on the same label for all categories. We used IBM SPSS Statistics 23 software to assess the level of agreement among the 3 experts for the morphological classification of the database. Statistical differences between experts in each morphology class were evaluated by Fisher’s exact test and considered significant at p < 0.05.

2.3. Deep learning approach

Our algorithm was developed using a convolutional neural network (CNN) architecture, which consists of five stages: image pre-processing, database partitioning, data augmentation, program training, and evaluation.

The algorithm was implemented in Python (version 3.8).

2.3.1. Image pre-processing

One of the main issues with the automated image-processing systems lies in the presence of noise signals that overlap with the sperm images. These signals can be attributed to insufficient lighting in optical microscopes or poorly stained semen smears [19]. The image-processing step aims to denoise the images, enabling an accurate estimation of each spermatozoon’s signal and improving classification by the model.

The data processing steps before creating a classification system typically involve several crucial stages to prepare and transform raw data into a suitable format for training a machine learning (ML) model. Here are the steps followed in our case:

1. Data Cleaning: Identify and handle missing values, outliers, or any inconsistencies in the dataset. Cleaning the data ensures that the model is not influenced by noise or inaccuracies that might hinder its performance.

2. Normalization/Standardization: Normalize or standardize numerical features to bring them to a common scale. This ensures that no particular feature dominates the learning process due to differences in magnitude. In our case, we resized images with linear interpolation strategy to 80*80*1 grayscale.

2.3.2. Partitioning

Splitting the Data: The entire set of images was divided into two subsets randomly: 80% of the dataset was selected to train the model, and the rest (20%) was used to test it. The training set is used to train the model, while the testing set is reserved to evaluate its performance on unseen data.

From the training subset we extract 20% of data as a validation set.

Handling Imbalanced Data: If the dataset has imbalances in class distribution, we consider applying techniques such as oversampling, to handle imbalanced datasets.

2.3.3. Data augmentation

Deep learning has demonstrated impressive outcomes in various computer vision tasks, often relying on substantial amounts of training data to prevent overfitting. However, the availability of labeled data for real-world applications can be limited. To address this challenge, DA has become an indispensable component in the training deep learning models with image data [20].

As an effective means to enhance the adequacy and diversity of training data, DA has emerged as an essential element in the successful application of deep learning models to image data [20].

DA techniques are applied to artificially increase the size of the training dataset and improve the model’s robustness. Various techniques of DA, such as horizontal flipping, rotation, and zooming, are available [17].

We employed the following DA techniques: horizontal mirroring, vertical mirroring, rotation at various angles, and auto-straightening to extend the dataset. Data augmentation was performed in Python using Keras library. These techniques were selected to preserve the original morphological class of each image.

All transformations were applied to the images. However, the number of augmentations generated per image was not equal across all classes. To reduce class imbalance, a higher number of augmentations was applied to rare morphological categories, whereas more common classes were augmented to a lesser extent.

2.3.4. Architecture of the proposed ML system

The neural network architecture is designed with meticulous consideration, employing the Adam optimizer for weight updates over a 100-epoch training process, utilizing a batch size of 32 samples per iteration. A learning rate of 0.001 is chosen to regulate the step size in weight updates, and a momentum of 0.99 is applied to efficiently navigate the parameter space. This configuration is meticulously crafted to ensure stable convergence and computational efficiency, reflecting a well-tuned combination of hyperparameters aimed at facilitating robust model training for sperm morphological classification.

The proposed Convolutional Neural Network (CNN) architecture is a key element in this design, incorporating a Convolutional 2D layer, Batch Normalization layer, Max Pooling layer, Flatten layer, and Dropout layer.

The Convolutional 2D layer is strategically designed to capture spatial hierarchies within the input data, enabling effective feature extraction through the application of convolutional filters. The introduction of Batch Normalization serves to normalize intermediate feature maps, mitigating internal covariate shift and promoting stable training. The Max Pooling layer plays a crucial role in down sampling spatial dimensions, extracting essential features while reducing computational complexity. The Flatten layer follows, reshaping the output into a one-dimensional vector, preparing it for input into fully connected layers. Finally, the Dropout layer is strategically placed to randomly deactivate 20% of neurons during training, acting as a preventative measure against overfitting and enhancing the model’s generalization capabilities. This combination of layers forms a well-structured convolutional neural network, adept at extracting hierarchical features, normalizing activations, down sampling spatial dimensions, and fostering robustness through dropout regularization. (refer to Figure 2 for an illustrative representation).

Figure 2.

CNN-based model for sperm morphological classification.

2.3.5. Evaluation

The evaluation of a ML algorithm constitutes a crucial aspect of any project. To assess the performance of our proposed classification model, various types of evaluation metrics were employed. Key metrics, including true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), have been acquired. These metrics provide a detailed understanding of the algorithm’s efficacy in correctly identifying positive and negative instances, facilitating a comprehensive evaluation of its overall performance and highlighting areas that may require further refinement or optimization.

Subsequently, we calculated Sensitivity (recall) (Se), Specificity (Sp), Positive predictive value (Precision), Accuracy and Kohen’s Kappa. The formulas for calculating these metrics are the following:

$$\mathit{\text{Accuracy}}=\frac{\mathit{\text{TP}}+\mathit{\text{TN}}}{\mathit{\text{TP}}+\mathit{\text{TN}}+\mathit{\text{FP}}+\mathit{\text{FN}}}$$

$$\mathit{\text{Recall}}=\frac{\mathit{\text{TP}}}{\mathit{\text{TP}}+\mathit{\text{FN}}}$$

$$\mathit{\text{Precision}}=\frac{\mathit{\text{TP}}}{\mathit{\text{TP}}+\mathit{\text{FP}}}$$

$$\mathit{\text{Specificity}}=\frac{\mathit{\text{TN}}}{\mathit{\text{TN}}+\mathit{\text{FP}}}$$

$$\mathit{\text{Kohen}}\text{'}s\hspace{0.17em}\mathit{\text{Kappa}}=\frac{2\ast \left(\mathit{\text{TP}}\ast \mathit{\text{TN}}-\mathit{\text{FN}}\ast \mathit{\text{FP}}\right)}{\left(\mathit{\text{TP}}+\mathit{\text{FP}}\right)\ast \left(\mathit{\text{FP}}+\mathit{\text{TN}}\right)+\left(\mathit{\text{TP}}+\mathit{\text{FN}}\right)\ast \left(\mathit{\text{FN}}+\mathit{\text{TN}}\right)}$$

These metrics provide a comprehensive assessment of the model’s performance, covering aspects such as its ability to correctly identify positive and negative instances, precision in positive predictions, overall accuracy, and the agreement beyond chance measured by Cohen’s Kappa. Incorporating these metrics in the evaluation process enhances the understanding of the model’s strengths and areas for potential improvement.

We also used a confusion matrix to evaluate the performance of the model. It displays the number of accurate and inaccurate instances based on the model’s predictions.

3. Results

3.1. Dataset description

The SMD/MSS dataset included 1000 images that had been categorized into one or more morphological classes by three experts using the modified David classification. Since there were multiple images with two labels or more, the total number of images categorized into the different classes was 1398 (Table 1).

Table 1.

The number of images before and after the data augmentation techniques.

Classes	Number of images before DA	Number of images after DA
Tapered head (a)	40	240
Thin head (b)	66	396
Microcephalous (c)	23	92
Macrocephalous (d)	14	84
Multiple heads (e)	12	72
Abnormal post-acrosome region (f)	479	1407
Abnormal acrosome (g)	343	1713
Cytoplasmic droplet (h)	78	390
Bent (j)	155	776
Short tail (l)	10	80
Coiled tail (n)	51	266
Multiple tails (o)	11	55
Normal (nr)	116	464
Total	1398	6035

The microscopic images depicted in Figure 3 showcase an example of sperm image from the dataset after the application of various DA techniques.

Figure 3.

Data augmentation by different techniques (a) original sperm image; (b) horizontal mirror; (c) vertical mirror; (d) auto-straightening (e) rotation 15°.

Following the application of DA techniques, the total number of images in the dataset was increased from 1398 to 6035 images. Table 1 provides a concise summary detailing the distribution of images across different sperm morphological classes both before and after application of DA.

3.2. Inter experts agreement

The level of agreement among experts was evaluated on dataset before DA, i.e., on 1000 images. The first set (scenario 1-NA) contains only 26 spermatozoa (2.6% of the dataset). The second set (scenario 2-PA) contains 318 spermatozoa, meaning that for 31.8% of the dataset, there was an agreement among 2/3 experts for at least one morphological class. The third set (scenario 3-TA) contains 656 spermatozoa (65.6% of the dataset), with total agreement between the three experts for all morphological classes.

Table 2 shows the morphological class distribution for each agreement scenario.

Table 2.

Number (%) of sperm per morphological class for the different agreement scenario.

Morphology class	Scenario 1 N1 = 26 (%)	Scenario 2 N2 = 318 (%)	Scenario 3 N3 = 656 (%)	Total N = 1000 (%)
a	4 (15.3%)	42 (13.2%)	26 (4%)	72 (7.2%)
b	2 (7.7%)	65 (20.4%)	52 (7.9%)	119 (11.9)
c	3 (1.2%)	11 (3.4%)	23 (3.5%)	37 (3.7%)
d	2 (7.7%)	11 (3.4%)	12 (1.8%)	25 (2.5%)
e	1 (3.8%)	11 (3.4%)	12 (1.83%)	24 (2.4%)
f	26 (100%)	265 (83.3%)	290 (4.42%)	581 (58.1%)
g	25 (96%)	265 (83.3%)	438 (66.8%)	728 (72.8%)
h	7 (27%)	37 (11.6%)	63 (9.6%)	107 (10.7%)
j	16 (61.5%)	143 (45%)	164 (25%)	323 (32.3%)
l	5 (19.2%)	12 (3.8%)	11 (1.7%)	28 (2.8%)
n	7 (27%)	41 (13%)	99 (15%)	147 (14.7%)
o	1 (3.8%)	7 (2.2%)	13 (2%)	21 (2.1%)
Nr	1 (3.8%)	39 (12.2%)	89 (13.6%)	149 (14.9%)

Scenario 1: No agreement among the experts; Scenario 2: partial agreement: 2/3 experts agree on the same label for at least one category, Scenario 3: total agreement: 3/3 experts agree on the same label for all categories.

In addition to our primary analyses, we conducted an in-depth investigation into inter-expert variability across various morphological classes, as detailed in Table 3. This examination aimed to provide a comprehensive understanding of the consistency or divergence in expert assessments within distinct morphological classes.

Table 3.

Variability among experts for the different morphological classes.

Classes	Number of spz for expert 1	Number of spz for expert 2	Number of spz for expert 3	Fisher’s exact test
A	60	63	53	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
B	79	87	101	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
C	32	34	25	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
D	22	20	18	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
E	21	23	21	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
F	600	588	639	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
G	684	702	647	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
H	96	101	103	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
J	269	372	285	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
L	21	11	11	p1 = 0.493
				p2 = 0.621
				p3 < 0.001**
N	148	150	135	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
O	16	17	19	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**
Nr	116	113	107	p1 < 0.001**
				p2 < 0.001**
				p3 < 0.001**

Spz: spermatozoa; p1: agreement between expert 1 and expert 2; p2: agreement between expert 1 and expert 3; p3: agreement between expert 2 and expert 3.

**p < 0.001.

The statistical analysis conducted using Fisher’s exact test revealed a significant overall agreement among the three experts (p < 0.001) for all morphological classes, except for class l. Notably, for class l there was no significant agreement between expert 1 and 2 (p = 0.493) and between expert 1 and 3 (p = 0.621).

3.3. Performance evaluation

The performance evaluation of our model has yielded comprehensive results, including confusion matrix (Figure 4) for different classes and precision statistics encompassing true positives, false positives, sensitivity, specificity, precision, and Cohen’s Kappa (Table 4). The confusion matrix showed the accuracy of the model’s classification on the dataset. Confusion was noticeable for between the class pairs (f,g) and (h,j). For the rest, agreement was the norm.

Figure 4.

Confusion matrix of the model.

Table 4.

The performance of the deep learning model for the different morphological classes.

Classes	Se	Sp	Precision	Accuracy	Cohen’s Kappa
A	43%	97%	71%	92%	0.6
B	57%	94%	50%	91%	0.48
C	5%	100%	100%	92%	0.9
D	36%	98%	90%	92%	0.23
E	12%	98%	50%	92%	0.7
F	61%	51%	43%	55%	0.3
G	64%	45%	56%	55%	0.9
H	44%	76%	45%	67%	0.8
J	83%	26%	62%	60%	0.9
L	79%	91%	86%	86%	0.71
N	82%	30%	61%	63%	0.23
O	17%	98%	60%	88%	0.3
Nr	62%	85%	69%	79%	0.52

In Table 4, we present a detailed overview of the deep learning model’s efficacy in sperm classification across various classes. Sensitivity ranged from 5% to 82%, with better performance for tail anomalies (l and n). Specificity ranged from 26% to 100%, with better performance for microcephalous (c). Precision ranged from 43% to 100%, with better performance for head anomalies (c, d, e) and multiple tails (o). Accuracy ranged from 55% to 92%, with better performance for head anomalies (a, c, d and e). Cohen’s Kappa ranged from 0.23 to 0.9, with better performance for microcephalous (c), abnormal acrosome (g) and bent spermatozoa (j).

4. Discussion

Sperm morphology assessment is of major importance during male infertility investigation. Traditionally, this assessment has relied on visual inspection, strongly linked to the expertise of the observer. The development of automated systems could therefore limit the variability linked to the human factor [6]. However, automated morphological classification of human sperm cells presents significant challenges, as demonstrated by the public SCIAN dataset [21]. These challenges can be attributed to several factors, including noisiness in low-magnification microscopic images of spermatozoa, insufficient examples of sperm cells for complex ML training, and the issue of class imbalance [22]. The automation of sperm morphology analysis has been attempted in several studies. Nevertheless, despite the development of several AI-based algorithms for classifying sperm morphology, published databases remain very scarce. All of these databases had a relatively limited number of images. In fact, the database with the largest number of images is that of Chang et al. (1854 images) [21]. Furthermore, the majority of these databases only focus on sperm head morphology and the classification have been carried out by a single expert. There are also a few ML approaches associated with the morphological classification of human sperm heads. Furthermore, it is important to note that there is currently no publicly available database categorized according to the modified David classification. All publicly available databases published to date are classified according to WHO classification [15] or Kruger classification (strict WHO 2010 criteria).

This study introduces a fully automated approach to sperm morphology analysis using the dataset SMD/MSS created in our laboratory. Comparing this database with others already published [19,21–30], we can identify several strengths, which enhances its utility for improving current approaches to automated sperm morphological classification. Firstly, this is the only dataset classified according to the modified David classification. Second, it is the pioneering dataset that associates each image with a ground truth file containing relevant information about the image. Furthermore, for each spermatozoon, the width and length of the head, as well as the length of the tail, were accurately determined using the CASA morphometric tool. This enhanced the precision and, consequently, the performance of our dataset for training the CNN-based model for size anomalies (classes a, b, c, d, and l). Moreover, our dataset included 13 morphological classes, representing the largest number of anomalies compared to other studies [19,21–30].

SMD/MSS dataset is among the few that have been classified by more than one expert. The only dataset labeled by three domain experts is the SCIAN-MorphoSpermGS database [21]. The quality of dataset is crucial to achieving high performance in machine learning. The level of agreement was excellent among our experts and higher than that reported by Chang and all [21] (TA 65.5% vs 20% and NA 2.6% vs 19%, respectively). The agreement was very high for almost all categories (Table 2). These results underscore the consistency among the experts in their assessments, reinforcing the reliability of the observed agreements across various morphological classes, with the exception of class l (short tail). The observed discrepancies for short tail anomaly have no impact on the performance of our AI-based system since they concern the evaluation by the experts (manual method). For the training of the system, we used the precise morphometric measurements carried out by the CASA system, which represents a good methodological choice in our study.

The size of the dataset is one of the most critical determinants of model performance. The number of images in our database is significant compared to other databases after the use of DA techniques. For data augmentation, we employed horizontal mirroring, vertical mirroring, rotation at various angles, and auto-straightening. The number of images in our dataset reached 6035 after DA techniques, surpassing that of other published datasets: 1457 (HSMA-DS) [26], 1854 (SCIAN-MorphoSpermGS) [21], 216 (HuSHeM) [26], 1540 (MHSMA) [27], and 200 (SMIDS-morpho) [19]. This augmentation process not only expanded the dataset but also contributed to a more comprehensive representation of various morphological features, enhancing the robustness and diversity of the training set for the subsequent phases of model development and evaluation.

The second aspect of our work involved developing a CNN model for sperm morphology classification. Our analysis showed that the model achieved an overall accuracy rate of 77.8% with sensitivity percentages ranging from 5% for the class “c” to 83% for class “j”, specificity percentages ranging from 26% (class “j”) to 100% (class “c”), precision percentages ranging from 43% (class “f”) to 100% (class “c”), accuracy percentages ranging from 55% (classes “f” and “g”) to 92% (classes “a”, “c”, “d” and “e”) and Cohen’s Kappa equal to 0.9 for classes “c”, “g” and “j” (Table 4). The results of our proposed model are satisfactory compared to previously published models [19,21–30] (Table 5).

Table 5.

Performance of various algorithms for automated classification of sperm morphology in the literature.

Study	Data set information	Accuracy (%)
Yi et al. 1998 [23]	300–360 images	58.2 to 96.7
Ramos et al. 2002 [24]	590 images	70
Ghasemian et al. 2015 [25]	HSMA-DS: 1457 images	88.05 to 93.73
Chang et al. 2017 [21]	SCIAN-MorphoSpermGS	33 to 49
	1854 images
Shaker et al. 2017 [26]	SCIAN-MorphoSpermGS	78,9 to 92,9
	1854 images
	HuSHeM	58 to 62
	216 images
Javadi et al. 2019 [27]	MHSMA: 1540 images	67.67 to 91.33
Riodan et al. 2019 [28]	SCIAN-MorphoSpermGS	62
	HuSHeM	62 to 72
Iqbal et al. 2019 [22]	SCIAN-MorphoSpermGS	88
	HuSHeM	95
Chandra et al. 2022 [29]	MHSMA	71 to 87.33
Soares et al. 2022 [30]	SCIAN-MorphoSpermGS	70
	HuSHeM	97,7
Ilhan et al. 2022 [19]	SMIDS	71.4 to 78.4
	200 images	62.8 to 75.8
	HuSHeM
Our study	SMD/MSS	55 to 92

Populating the dataset with more images in a targeted manner would have helped improving the performance of the model.

It is worth noting the limitations of this study. Although our laboratory is a reference center and our experts have good experience, it would be interesting to include the evaluation of other experts in our dataset to reach a consensus on sperm classification. Additionally, our model is created based on images of a specific quality, which needs to be verified on images with other processing tools or using other staining methods.

Our dataset provides a crucial foundation for advancing research in various directions: testing other classification models and combining the results of several models as previously recommended [21], using the WHO classification to reach more users worldwide, and developing a functional classification model for AI-based motile sperm selection during intracytoplasmic sperm injection (ICSI).

5. Conclusion

Sperm morphology analysis is considered an important marker of male fertility. The traditional manual assessment is time-consuming and relies on professional expertise, leading to significant intra and inter-laboratory variability. To address these issues, we proposed a CNN-based model to automate the morphological analysis of human spermatozoa. This CNN model was created, and from a computational perspective, its performance was primarily dependent on the robustness of the datasets used for training. Until now, there have been few publicly available sperm morphology datasets. Therefore, we used SMD/MSS dataset consisting of 1000 images, obtained from human sperm smears prepared according to the 2021 WHO recommendations. Three experts carried out the morphological classification of sperm according to the modified David classification with a good level of agreement. Various DA techniques were applied to enhance the dataset size, multiplying it sixfold, and thus achieving better performance. The second contribution of this study is the validation of our automated CNN model for the morphological classification of spermatozoa. The results indicated the excellent performance of the model in predicting different morphological classes. In fact, our proposed CNN-based model achieved 77.8% of overall accuracy rate.

The SMD/MSS dataset and the related classification model are intended to become a reference and a foundation for future improvements in current approaches to human spermatozoa classification. Our dataset represents an important basis for developing research in several directions: testing and combining other classification models, using the WHO classification in order to reach more users around the world, and developing a functional classification model for AI-based motile sperm selection during intracytoplasmic sperm injection (ICSI).

Funding Statement

This paper was not funded.

Ethics approval and consent to participate

All participants signed a consent form, and we obtained approval from the Committee for the Protection of Persons (CPP) of South Tunisia. The approval number is 0537/2023.

Authors contributions

Olfa Abdelkefi: data curation, formal analysis, software, writing and original draft. Rania Maalej: methodology, software, writing and editing. Tarek Rebai: Review and editing. Afifa Sellami: Review and editing. Salima Daoud: conceptualization, supervision, methodology and writing.

Disclosure statement

The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Data availability statement

Data will be made available on request. (https://drive.google.com/drive/folders/1EUNxTSeFRju3xAVeqQ1MzSSLKidP6wzD?usp=sharing).