A method and system for classifying sperm heads in the presence of mixed noise

Abstract

The application provides a sperm head classification method and system for mixed noise, and the method comprises the following steps: constructing a sperm head image classification model combining denoising and classification; obtaining a sperm head image dataset containing annotation information of multiple annotation sources; taking each annotation source as an independent learning task to perform multi-task learning training; inputting a sperm head image to be evaluated into the trained sperm head image classification model to obtain a sperm head classification result. The application adopts a multi-task learning mode, inputs the annotation information of different annotation sources into the model as independent tasks for training, and performs weighted processing on the samples according to the uncertainty and category frequency distribution of the samples, so that the interference of high-uncertainty samples on the model training is weakened while all the annotation information is retained, thereby significantly improving the accuracy and robustness of the sperm head classification model in the presence of label noise and image noise.

Description

Technical Field

[0001] This invention relates to the field of sperm image processing technology, specifically to a method and system for classifying sperm heads in the context of mixed noise. Background Technology

[0002] Over the past few decades, the use of assisted reproductive technology (ART) has increased significantly. However, the success rate of ART has remained at around 50% per cycle. Therefore, rigorous screening and evaluation of sperm quality are essential during ART. The main types of abnormal sperm defects include: head, neck, tail, and excessive residual cytoplasm. Abnormalities in the sperm neck and tail are highly subjective and difficult for physicians to identify, and are not the most common or critical characteristic for determining sperm morphological abnormalities. In contrast, sperm head examination has become the primary basis for assessing sperm morphological abnormalities in domestic and international research and clinical practice.

[0003] Early methods in sperm head morphology classification studies were largely based on geometric feature extraction. For example, sperm morphology features were divided into two categories: the first category included dimensional features such as length, width, and area; the second category included proportional or morphological features such as aspect ratio, length-to-width difference, ellipticity, and shape. By analyzing a large amount of sperm head morphology data, the correlations between different categories of sperm head morphology were summarized. Later, machine learning methods were gradually introduced into this field, including Bayesian density estimation, Otsu thresholding, and support vector machines. In recent years, with the development of deep learning, deep neural network structures such as convolutional neural networks have also been widely applied to the automatic classification of sperm head morphology, achieving good results.

[0004] However, the number of expert labelers in the field of sperm is limited, and obtaining high-quality, multi-expert labeled data is extremely costly. Existing sperm head image labeling is typically done by a small number of experts, and different experts may produce inconsistent labels for the same image. For example, in the SCIAN dataset, although three experts participated in the labeling, the data was only retained when two or more experts agreed, and the maximum voting rule was used to determine the label; otherwise, the sample was discarded. This approach results in a large amount of labeled information being unused, and the labels obtained by the maximum voting rule may themselves be incorrect, thus introducing label noise.

[0005] Furthermore, during image acquisition, sperm head images are prone to degradation due to factors such as continuous sperm movement, focal plane drift, and changes in liquid refraction, leading to weakened structural information and increased background noise. Existing methods typically treat image denoising and morphological classification as two independent tasks, failing to fully explore the intrinsic relationship between them and limiting further improvements in classification performance. Summary of the Invention

[0006] This invention proposes a sperm head classification method and system for mixed noise, which overcomes the interference of label noise and mixed noise in sperm images caused by differences in multi-expert annotation, and improves the robustness of head morphology classification.

[0007] To address the aforementioned technical problems, this invention provides a method for classifying sperm heads in the presence of mixed noise, characterized in that the method includes:

[0008] Step S1: Construct a sperm head image classification model that combines denoising and classification. The sperm head image classification model includes an image denoising module, a feature extraction module, and a classification module connected in sequence.

[0009] Step S2: Obtain annotation information from multiple annotation sources, where each annotation source represents a annotation perspective for classifying sperm heads; treat the annotation information from each annotation source as an independent learning task and perform multi-task learning training on the sperm head image classification model; the total loss function used in the multi-task learning training includes at least a multi-task classification loss, which is used to minimize the weighted cross-entropy loss between the model prediction results corresponding to each independent task and their corresponding annotation source annotation information;

[0010] Step S3: Input the sperm head image to be evaluated into the trained sperm head image classification model to obtain the prediction results corresponding to each independent task. Combine the prediction results of each independent task to obtain the final sperm head classification result.

[0011] Preferably, in the sperm head image classification model, the image denoising module converts the input image to a frequency domain representation via Fourier transform; uses a frequency domain filter to modulate the spectral amplitude to suppress high-frequency interference and retain mid-to-low frequency structural information; and restores it to the spatial domain via inverse Fourier transform to enhance edge contours and detailed features.

[0012] This denoising process effectively separates noise from valid signals through frequency domain filtering. While suppressing high-frequency interference noise, it retains mid-to-low frequency structural information (such as the outline of the sperm head). After inverse transformation, it significantly improves the image signal-to-noise ratio, enhances edge sharpness and texture details, and provides clearer morphological features for subsequent classification.

[0013] Preferably, the sperm head image classification model in step S1 includes a saliency modeling module between the image denoising module and the feature extraction module. The saliency modeling module is used to divide the image after processing by the image denoising module into several non-overlapping image blocks, calculate the saliency score of each image block through a shallow neural network, and use the saliency score as a weight to perform weighted modulation on the embedded features, so as to enhance the attention to the sperm head region and suppress background interference.

[0014] Preferably, the feature extraction module adopts a VMamba-based network architecture, which includes multiple cascaded VSS modules. Each VSS module includes spatially selective 3D operations, depthwise separable convolutions, and residual connections. This architecture captures global morphological features through spatially selective 3D operations, improves parameter efficiency through depthwise separable convolutions, and enhances gradient propagation through residual connections, achieving large receptive field modeling and significantly improving the ability to recognize subtle morphological differences in sperm heads (such as blurred boundaries and texture variations).

[0015] Preferably, the multi-task classification loss The expression is:

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] In the formula, This indicates the total number of sperm samples. This indicates the total number of sperm head types; Indicates the number of labeled sources; Indicates the first The first annotation source in the first Sample weights on each sample; Indicates the first The first annotation source in the first Labels for each sample The value is 1 if it is c, otherwise it is 0. This represents the predicted probability of the sample. Indicates the first One sample, This represents the parameter corresponding to the j-th independent branch of the annotation source; express Frequency of the corresponding category; Indicates the first Inconsistency among individual samples; Indicates the first The frequency distribution of samples labeled as class c; Indicates the penalty coefficient; Indicates the first The number of samples labeled as class c.

[0021] Preferably, the total loss function further includes parameter consistency loss. The expression is:

[0022] ;

[0023] In the formula, This represents the Frobenius norm.

[0024] The parameter consistency loss forces the model to learn common features between different annotation sources by constraining the similarity of network parameters in different task branches. This suppresses the annotation perspective bias caused by overfitting to a single annotation source, improves the model's generalization ability, and ultimately enhances the robustness and reliability of the classification results.

[0025] Preferably, the total loss function further includes total variation loss. The expression is:

[0026] ;

[0027] In the formula, Represents total variation in image processing; Indicates sample The output after processing by the image denoising module.

[0028] Total variation loss effectively suppresses noise interference by constraining the total variation value output by the image denoising module, while preserving key structural features of the sperm head (such as edges and textures), significantly improving the image signal-to-noise ratio and providing clearer and more stable input features for subsequent classification tasks.

[0029] Preferably, the sperm head image classification model is trained using an alternating update method: the parameters of the classification module are updated while the parameters of the image denoising module are fixed, and then the parameters of the image denoising module are updated while the parameters of the classification module are fixed again, until the network converges.

[0030] This invention also provides a sperm head classification system for mixed noise, implemented based on the above method, comprising:

[0031] Image denoising module: It is used to convert the input sperm head image to the frequency domain representation through Fourier transform, use the frequency domain filter to modulate the spectrum amplitude to suppress high frequency interference and retain mid- and low frequency structural information, and then use inverse Fourier transform to restore it to the spatial domain to enhance edge contours and detailed features.

[0032] Feature extraction module: connected to the image denoising module, used to extract global features of the denoised image, and introduce spatially selective 3D operations and depthwise separable convolution to construct a spatial modeling path with a large receptive field to adapt to the slight differences in shape, boundary and structural texture of different sperm head morphologies;

[0033] The classification module is connected to the feature extraction module and includes multiple independent branches of annotation sources. Each independent branch of annotation source corresponds to the annotation information of an annotation source. The prediction results of each independent task are fused to obtain the final sperm head classification result.

[0034] Preferably, a saliency modeling module is provided between the image denoising module and the feature extraction module. The saliency modeling module is used to divide the denoised image into several non-overlapping image blocks, calculate the saliency score of each image block through a shallow neural network, and use the saliency score as the weight to perform weighted modulation on the embedded features, so as to enhance the attention to the sperm head region and suppress background interference.

[0035] The beneficial effects of the present invention include at least the following:

[0036] 1) To address the problem of high cost and inconsistent results in sperm image annotation by different experts, this invention adopts a multi-task learning approach, using annotation information from different annotation sources as independent task inputs to train the model, and weighting the samples according to the uncertainty and class frequency distribution. This approach can retain all annotation information while weakening the interference of high uncertainty samples on model training, thereby significantly improving the accuracy and robustness of the sperm head classification model in the presence of label noise.

[0037] 2) To address the problem that existing methods separate sperm head image denoising and classification tasks and fail to fully utilize their inherent correlation, this invention adopts an end-to-end learning approach that combines denoising and classification. This allows the denoising module and the classification module to promote each other during training and fully explore the global features of the sperm image during the feature extraction stage. This achieves implicit denoising of the sperm image, which not only effectively suppresses background interference but also further improves the overall performance of sperm head classification.

[0038] 3) Through the dual optimization of frequency domain denoising module and image patch saliency weighting mechanism, noise interference is effectively suppressed in the preprocessing stage, and key areas of sperm head are focused in the feature extraction stage, which significantly improves the ability to identify and classify difficult samples with blurred boundaries. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the image denoising module and the classification joint network used in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of a sperm head image and its multiple expert annotations, representing an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0043] like Figure 1 As shown, this embodiment of the invention provides a method for sperm head classification in the face of mixed noise, including:

[0044] Step S1: Construct a sperm head image classification model that combines denoising and classification. The sperm head image classification model includes an image denoising module, a feature extraction module, and a classification module connected in sequence.

[0045] Step S2: Obtain a sperm head image dataset containing annotation information from multiple annotation sources, where the annotation information from each annotation source represents an annotation perspective; treat the annotation information from each annotation source as an independent learning task, and use the dataset to perform multi-task learning training on the sperm head image classification model; the total loss function used in the multi-task learning training includes at least a multi-task classification loss, which is used to minimize the weighted cross-entropy loss between the model prediction results corresponding to each independent task and their corresponding annotation information from the annotation source;

[0046] Step S3: Input the sperm head image to be evaluated into the trained sperm head image classification model to obtain the prediction results corresponding to each independent task. Combine the prediction results of each independent task to obtain the final sperm head classification result.

[0047] Specifically, in step S1, an end-to-end model for simultaneous denoising and classification of sperm head images is constructed, and the global information of the image is fully explored in the network design, such as... Figure 2 As shown, to improve the accuracy and robustness of human sperm head morphology classification, this invention proposes an end-to-end deep neural network architecture for joint denoising and classification. The overall network consists of three main modules: an image denoising module, a feature extraction module, and a classification module. In this embodiment, the image denoising module can employ, but is not limited to, mean filtering, median filtering, bilateral filtering, low-pass filtering, wavelet transform denoising, denoising methods based on convolutional neural networks, and denoising methods based on learnable frequency domain filters. The following description uses a frequency domain-based image denoising module.

[0048] During image acquisition, various factors such as the continuous movement of sperm in the liquid culture environment, focal plane drift, and changes in liquid refraction often result in weakened structural information and enhanced background noise, leading to significant overall quality degradation. These imaging defects severely affect the model's ability to perceive key regions, limiting the performance of sperm morphology classification tasks. To alleviate this problem, a frequency-domain-based image denoising module is introduced at the model's front end. This module first converts the image to a frequency domain representation through Fourier transform and modulates the spectral amplitude using a matrix-based frequency domain filter, effectively suppressing high-frequency interference components and background noise while preserving structural information in the mid- and low-frequency regions. Subsequently, an inverse Fourier transform is used to restore the image to the spatial domain, significantly enhancing edge contours and morphological details, providing a higher-quality input foundation for subsequent feature modeling.

[0049] After denoising, the sperm image is divided into several fixed-size, non-overlapping image patches, which are then embedded into a high-dimensional feature space through linear transformation to construct the input sequence for the feature extraction module. To further enhance the model's ability to focus on key regions of the sperm head, this embodiment introduces an image patch-level saliency modeling mechanism. This mechanism automatically learns the saliency score of each image patch through a shallow neural network and uses this score as a weight to weight and modulate the embedded features, guiding the model to focus on the structural features of the sperm head region. This effectively suppresses the interference of redundant background regions on the overall discrimination performance, especially demonstrating stronger discriminative ability when dealing with categories with blurred boundaries such as small-headed and cone-shaped sperm heads.

[0050] In this embodiment, the feature extraction backbone adopts a VMamba-based network architecture. This architecture uses multi-layered, cascaded VSS modules as its basic building blocks, possessing excellent cross-spatial modeling capabilities and parameter efficiency. Each VSS module introduces spatially selective 3D operations and depthwise separable convolutions to construct a spatial modeling path with a large receptive field, adapting to subtle differences in shape, boundary, and structural texture among different sperm head morphologies. The residual connection mechanism within the modules further enhances the stability of feature representation and gradient propagation capabilities, improving the model's recognition performance on difficult-to-classify samples, such as structurally similar but vaguely defined categories.

[0051] During the classification phase, the model normalizes, pools, and flattens the extracted high-dimensional global features, and finally outputs five sperm morphology labels through a fully connected layer. The overall architecture supports end-to-end training and exhibits excellent discrimination performance under complex imaging conditions.

[0052] To address the problem of label noise caused by differences in annotation information from different sources, this invention proposes a concept that utilizes the differentiated standard information itself as the basis for multi-task learning. In step S2, the constructed image dataset contains multiple annotation sources that may have differing annotation information, such as experts, algorithms, different annotation standards, different databases, and any other source that generates annotations.

[0053] Specifically, this embodiment of the invention collects the SCIAN dataset for sperm head classification, including different sperm head morphologies: normal head, cone-shaped head, pear-shaped head, round head, and small head. Figure 1 The image shows schematic diagrams of different sperm head morphologies in the SCIAN dataset, with annotations provided by three experts. Existing methods retain data if two or more experts agree on the annotations, using a maximum voting rule for final labeling. Otherwise, the annotated data is discarded. However, this approach fails to fully utilize the labeling information of some data, and the maximum voting rule may result in incorrect labels, leading to noisy labels. Furthermore, due to the presence of data noise, sperm head images need to be denoised before classification. Current methods treat sperm head image denoising and classification as two independent tasks, failing to fully explore the inherent correlation between them. Specifically, inconsistencies in annotation information from multiple different sources lead to label noise. How to fully utilize the information from different annotation sources to improve the accuracy of sperm head classification models under noisy labels is one of the key challenges in predicting whether sperm morphology is normal.

[0054] Therefore, in this embodiment, the annotation information of each annotation source (expert) is treated as a learning task, a loss function based on multi-task learning is constructed, and the parameters of the classification model are updated according to the loss function.

[0055] Specifically, assume there are N samples in the training set. Each sample has annotation information from M annotation sources. Unlike previous methods that simply used the maximum voting rule to determine the final label from multiple annotation sources for training, this invention fully utilizes the annotation information from each annotation source, treating each source's annotation information as a learning task. This allows for multi-task learning of the annotation information from multiple sources, ensuring full utilization of each source's annotation information while maintaining a consistent network structure for each learning task. Furthermore, annotations from different sources for each sample may be inconsistent; therefore, it is necessary to measure the inconsistency of the first annotation source. Sample Label Inconsistency ,remember The samples labeled as class c are: If there are 1, then the sample The frequency distribution is as follows:

[0056] ;

[0057] in This represents the total number of sperm head types. Therefore, the inconsistency in the labeling of each sample is calculated as follows:

[0058] ;

[0059] Where m is the penalty coefficient, set to a positive number greater than or equal to 2. The larger m is, the stronger the penalty for inconsistency in the annotation, and the greater the resulting inconsistency. The larger the value, the smaller the weight should be assigned to the corresponding sample when performing multi-task learning.

[0060] remember The frequency of the corresponding category is The corresponding sample weights are calculated as follows:

[0061] ;

[0062] Therefore, we weight the samples in multi-task learning. The lower the uncertainty of a sample and the higher the frequency of its corresponding class, the larger the weight. Otherwise, the weight is smaller. Thus, the corresponding loss function for multi-task learning is as follows:

[0063] ;

[0064] in This represents the parameter corresponding to the j-th independent branch of the labeled source in the learning network; express The value is 1 if it is c, otherwise it is 0. This represents the predicted probability of a sample.

[0065] To ensure that the labels predicted by each annotation source are as consistent as possible, the network parameters learned by each annotation source need to be as consistent as possible. Therefore, the corresponding parameter consistency loss function is as follows:

[0066] ;

[0067] In the formula, This represents the Frobenius norm.

[0068] In addition, such as Figure 3 As shown, the sample The output after the learnable Fourier transform is denoted as... For real, noisy sperm head images, it is impossible to obtain truly clean images for denoising. To achieve implicit denoising of sperm head images in a joint end-to-end denoising and classification network, this invention adopts the following total variation-based loss function:

[0069] ;

[0070] in The total variation in image processing is represented by the total variation. Since the presence of noise generally increases the total variation of an image, this embodiment of the invention minimizes the total variation of the output after the learnable Fourier transform, thereby achieving implicit image denoising.

[0071] In summary, we obtain the final total loss function. as follows:

[0072] ;

[0073] in and is the regularization coefficient.

[0074] This invention proposes a sperm head classification model that can simultaneously take into account both data noise and label noise present in sperm head classification, and fully exploit the global information of the image, thereby improving the accuracy and robustness of the sperm head classification model under mixed noise conditions.

[0075] In step S3, the sperm head image to be evaluated is input into the trained sperm head image classification model, the prediction results of each independent task are fused, and the final sperm head classification result is obtained by using the maximum voting rule.

[0076] Specifically, in addition to conventional training methods, this embodiment of the invention employs an alternating direction parameter training method. The learning rate is set to... The batch size is 32, and the training epochs are 500. Since the total variation loss function only contains a part of the network parameters, this invention uses an alternating update method to train the network parameters of the denoising part and the network parameters of the classification part. That is, first fix the parameters of the denoising part and update the parameters of the classification network, then fix the parameters of the classification network and update the parameters of the denoising part again, until the network finally converges.

[0077] Experiments were then conducted using the SCIAN dataset for sperm head classification. The classification performance of the method in this embodiment was compared with several typical networks. CNN and Transformer methods achieved accuracies of 77.1% and 82.3%, respectively, while the method in this embodiment achieved an accuracy of 88.7%, indicating that the overall performance of this embodiment is superior to existing methods. Furthermore, ablation experiments were conducted. Removing the image denoising module resulted in an accuracy of 87.5%, demonstrating that adding a denoising module improves the model's classification performance. Removing the consistency regularization loss resulted in an accuracy of 87%, indicating that ensuring the parameters learned from different annotation sources are as consistent as possible helps improve performance. In conclusion, extensive experiments demonstrate that the method proposed in this invention has superior performance.

[0078] This invention also provides a sperm head classification system for mixed noise, implemented based on the above method, comprising:

[0079] Image denoising module: It is used to convert the input sperm head image to the frequency domain representation through Fourier transform, use the frequency domain filter to modulate the spectrum amplitude to suppress high frequency interference and retain mid- and low frequency structural information, and then use inverse Fourier transform to restore it to the spatial domain to enhance edge contours and detailed features.

[0080] Feature extraction module: Connected to the image denoising module, it extracts global features from the denoised image and introduces spatially selective 3D operations and depthwise separable convolution to construct a spatial modeling path with a large receptive field to accommodate subtle differences in shape, boundary, and structural texture of different sperm head morphologies. A saliency modeling module is set between the image denoising module and the feature extraction module. The saliency modeling module is used to divide the denoised image into several non-overlapping image blocks, calculate the saliency score of each image block through a shallow neural network, and use the saliency score as a weight to perform weighted modulation on the embedded features to enhance the focus on the sperm head region and suppress background interference.

[0081] Classification module: Connected to the feature extraction module, it includes multiple independent tasks. Each annotation source branch corresponds to the annotation information of one annotation source. The prediction results of each annotation source branch are fused, and the final sperm head classification result is output based on the maximum voting rule.

[0082] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0083] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for classifying sperm heads in the presence of mixed noise, the method comprising: The method comprises: ​ Step S1: constructing a sperm head image classification model combining denoising and classification, the sperm head image classification model comprising an image denoising module, a feature extraction module and a classification module connected in sequence; Step S2: obtaining annotation information containing multiple annotation sources, the annotation information of each annotation source representing one annotation perspective of classifying sperm heads; taking the annotation information of each annotation source as an independent learning task, performing multi-task learning training on the sperm head image classification model; the total loss function used in the multi-task learning training at least comprises a multi-task classification loss, the multi-task classification loss being used for minimizing the weighted cross-entropy loss between the model prediction result corresponding to each independent task and the annotation information of the corresponding annotation source; Step S3: inputting a sperm head image to be evaluated into the trained sperm head image classification model to obtain prediction results corresponding to each independent task, fusing the prediction results of the independent tasks to obtain a final sperm head classification result; In step S1, the sperm head image classification model is provided with a saliency modeling module between the image denoising module and the feature extraction module, the saliency modeling module being used for dividing the image processed by the image denoising module into a plurality of non-overlapping image blocks, calculating the saliency scores of the image blocks through a shallow neural network, and weighting and modulating the embedded features by using the saliency scores as weights, so as to enhance the attention to the sperm head region and suppress background interference; The feature extraction module adopts a network architecture based on VMamba, the network architecture of VMamba comprising a plurality of VSS modules connected in series, each VSS module comprising a spatially selective three-dimensional operation, a depth separable convolution and a residual connection mechanism; The multi-task classification loss The expression is: ； ； ； ； In the formula, This indicates the total number of sperm samples. This indicates the total number of sperm head types; Indicates the number of labeled sources; Indicates the first The first annotation source in the first Sample weights on each sample; Indicates the first The first annotation source in the first Labels for each sample The value is 1 if it is c, otherwise it is 0. This represents the predicted probability of the sample. Indicates the first One sample, This represents the parameter corresponding to the j-th independent branch of the annotation source; express Frequency of the corresponding category; Indicates the first Inconsistency among individual samples; Indicates the first The frequency distribution of samples labeled as class c; Indicates the penalty coefficient; Indicates the first The number of samples labeled as class c.

2. The method for classifying sperm head towards mixed noise according to claim 1, characterized in that: In the sperm head image classification model, the image denoising module converts the input image into a frequency domain representation through Fourier transform; The frequency domain filter is used to modulate the spectral amplitude to suppress high-frequency interference and retain low-frequency structural information; The inverse Fourier transform is used to restore to the spatial domain to enhance the edge profile and detail features.

3. The method of claim 1, wherein: The total loss function further comprises a parameter consistency loss , the expression is: ； wherein denotes the Frobenius norm.

4. The method of classifying sperm heads according to claim 3, wherein: The total loss function further comprises a total variation loss , expressed as: ； In the formula, denotes a total variation in image processing; denotes a sample the output after processing by the image denoising module.

5. The method of claim 1, wherein: When training the sperm head image classification model, the training is performed in an alternating updating manner: the parameters of the classification module are updated while the parameters of the image denoising module are fixed, and then the parameters of the image denoising module are updated while the parameters of the classification module are fixed, until the network converges.

6. A hybrid noise oriented sperm head classification system, realized based on the method according to any one of claims 1 to 5, characterized in that, It comprises: An image denoising module: used for converting the input sperm head image into a frequency domain representation through Fourier transform, modulating the spectral amplitude by using a frequency domain filter to suppress high-frequency interference and retain low-frequency structural information, and then restoring to the spatial domain through inverse Fourier transform to enhance the edge profile and detail features; A feature extraction module: connected with the image denoising module, used for extracting the global features of the denoised image, introducing spatially selective three-dimensional operation and depth separable convolution, and constructing a spatial modeling path with a large receptive field to adapt to the slight differences in shape, boundary and structural texture of different sperm head morphologies; The classification module is connected with the feature extraction module and comprises a plurality of label source independent branches, each of which corresponds to label information of a label source, fuses prediction results of the independent tasks, and obtains a final sperm head classification result.

7. The system for classifying sperm heads in the presence of mixed noise of claim 6, wherein: The saliency modeling module is arranged between the image denoising module and the feature extraction module, and is configured to divide the denoised image into a plurality of non-overlapping image blocks, calculate saliency scores of the image blocks through a shallow neural network, and perform weighted modulation on the embedded features by taking the saliency scores as weights, so as to enhance the attention degree to the sperm head region and suppress background interference.

Images