High-precision sperm morphology screening system and screening method

Abstract

The invention discloses a high-precision sperm morphology screening system and screening method, and relates to the technical field of sperm cell recognition. Obtaining original semen and filtering to obtain initial semen; paving the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer motile sperms; performing a first amplification operation on the initial single-layer moving sperm to obtain a first initial single-layer moving sperm image; performing second amplification operation on the first initial single-layer moving sperm image to obtain a second initial single-layer moving sperm image; and substituting the second initial single-layer moving sperm image into a high-precision sperm morphology selection model to obtain a target high-quality sperm. According to the sperm morphological screening method, primary filtration of original sperm, secondary screening of a preset single-layer sperm strip laying method and a progressive amplification strategy of the first amplification operation and the second amplification operation are adopted, so that the problem that high-precision sperm morphological screening can be performed when high-quality sperms are selected by the existing IMSI (International Mobile Subscriber Identity) technology is solved, and the screening efficiency of the high-precision sperms is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sperm cell recognition, and particularly relates to a high-precision sperm morphology screening system and a screening method. BACKGROUND

[0002] Intracytoplasmic sperm injection (ICSI) is a common assisted reproductive technology for treating male infertility at present, and is suitable for patients with abnormal sperm morphology, low sperm quantity and low sperm activity. However, clinical data shows that the proportion of abnormal embryonic development of patients with severe teratospermia is still significantly higher than that of infertile men with normal sperm morphology even if the patients obtain embryos through the conventional ICSI technology, and the patients often show repeated pregnancy failure, which brings heavy physical and mental burden to the patient families. One of the important reasons is that the quality of sperm used for microinjection is poor. At present, the magnification of the microscope used in the ICSI process at home and abroad is about 200-400 times, and only sperm with no obvious abnormal morphology can be screened for injection, but it is difficult to identify the subtle structural abnormalities of the sperm head, such as vacuole and mitochondria abnormalities.

[0003] Medical Research and Education discloses intracytoplasmic morphologically selected sperm injection (IMSI) technology and sperm selection, and selecting sperm with excellent morphology for assisted reproductive treatment is an effective way to improve the clinical outcome. Normal sperm selected by a low magnification optical microscope (x200) may contain some subtle structural defects, and the sperm is used for intracytoplasmic sperm injection, which sometimes leads to fertilization failure or embryonic development arrest. Intracytoplasmic morphologically selected sperm injection technology is a perfect combination of ICSI and morphological examination of sperm organelles, and IMSI is equipped with differential interference phase contrast microscopy and digital imaging technology, so that the morphological quality of moving living sperm can be evaluated under high magnification (x6600), and sperm with no / less vacuoles in the head is selected for microinjection, so as to improve the fertilization rate, implantation rate and pregnancy rate and reduce the abortion rate. This paper mainly reviews the development of IMSI technology and its application in sperm selection and analyzes the clinical outcome of IMSI, so as to provide theoretical guidance for the application expansion and further function development of IMSI.

[0004] Patent CN109064469A discloses a sperm quality detector and a sperm quality detection system. The sperm quality detector comprises an image amplification device, an image acquisition device and an image processing device. The image amplification device is used for amplifying a sperm sample to be detected. The image acquisition device is used for acquiring an image of the amplified sperm sample, generating image information and sending the image information to the image processing device. The image processing device is used for analyzing and calculating the image information to generate a calculation result, and the calculation result includes sperm concentration and / or sperm activity.

[0005] However, the existing IMSI technology is only a simple judgment of the head of the sperm with few or no vacuoles, so that the high-precision sperm morphology screening is inefficient. SUMMARY

[0006] The purpose of the present application is to solve the problem of the existing IMSI technology in selecting high-quality sperm, which is only a simple judgment of the head of the sperm with few or no vacuoles, so that the high-precision sperm morphology screening is inefficient, and to propose a high-precision sperm morphology screening system and method.

[0007] In the first aspect of the present application, a high-precision sperm morphology screening method is first proposed, which comprises: Obtaining raw semen, filtering the raw semen to obtain initial semen; Paving the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm; Performing a first magnification operation on the initial single-layer active sperm to obtain a first initial single-layer active sperm image; Magnifying the first initial single-layer active sperm image to a target resolution to obtain a magnified initial single-layer active sperm image; Performing a second magnification operation on the first initial single-layer active sperm image according to the magnified initial single-layer active sperm image to obtain a second initial single-layer active sperm image; For the second initial single-layer active sperm image, a structure information image is obtained by a Sobel operator; Substituting the second initial single-layer active sperm image and the structure information image into a high-precision sperm morphology selection model to obtain target high-quality sperm.

[0008] Optionally, paving the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm comprises: Step 1, glass dish preparation: preparing N 20uL Gamete liquid strips and a drop of 2uL multi-edge antenna-like PVP liquid drop in the IMSI glass dish; Step 2, semen strip paving: for each Gamete liquid strip, 10uL of liquid in the Gamete liquid strip is sucked, and 10uL of initial semen is injected, and centrifugal semen paving is obtained by centrifugation for a predetermined period of time; Step 3, preliminary screening of sperm: under an inverted microscope, all initial normal sperm in the centrifugal semen paving are selected by a predetermined rule using an ICSI needle, and are added to the PVP liquid drop to obtain initial single-layer active sperm.

[0009] Optionally, performing a second magnification operation on the first initial single-layer active sperm image according to the magnified initial single-layer active sperm image to obtain a second initial single-layer active sperm image comprises: the first initial single-layer motile sperm image and the magnified initial single-layer motile sperm image into a pre-trained VAE encoder to obtain a low-resolution latent vector and a high-resolution latent vector; noise is added to the low-resolution latent vector and the high-resolution latent vector to obtain a low-resolution noise latent vector and a high-resolution noise latent vector; the high-resolution noise latent vector is updated according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector; a mean square error loss is generated according to the low-resolution latent vector and the updated high-resolution latent vector; the magnified initial single-layer motile sperm image is updated by gradient according to the mean square error loss to obtain a second initial single-layer motile sperm image.

[0010] Optionally, updating the high-resolution noise latent vector according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector comprises: a first initial single-layer motile sperm image is segmented by a preset first region size to obtain a set of labeled sub-initial single-layer motile sperm images, and a corresponding text prompt is generated for each sub-initial single-layer motile sperm image by a large language model; the low-resolution noise latent vector and the text prompt corresponding to each sub-initial single-layer motile sperm image are input into a preset Unet network to obtain a low-resolution attention score corresponding to each sub-initial single-layer motile sperm image; the high-resolution latent vector is segmented by a preset second region size to obtain a set of labeled sub-high-resolution latent vectors; the labels in the set of sub-high-resolution latent vectors correspond one-to-one to the labels in the set of sub-initial single-layer motile sperm images; the text prompt, the low-resolution attention score and the sub-high-resolution latent vector corresponding to the same label are input into an attention synthesizer for denoising to obtain a sub-updated high-resolution latent vector corresponding to the label; all sub-updated high-resolution latent vectors corresponding to the labels are merged to obtain an updated high-resolution latent vector.

[0011] Optionally, inputting the second initial single-layer motile sperm image and the structural information image into a high-precision sperm morphology selection model to obtain a target high-quality sperm comprises: the second initial single-layer motile sperm image and the structural information image are respectively input into a single residual block to obtain a basic feature and a structural basic feature; the basic feature and the structural basic feature are input into a multi-scale sampling model to obtain a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature; The low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature are substituted into a double-path feature optimization model to obtain a low-scale final fusion feature, a medium-scale final fusion feature and a high-scale final fusion feature; According to the low-scale final fusion feature, the medium-scale final fusion feature and the high-scale final fusion feature, a target high-quality sperm is screened in combination with a morphological quality standard.

[0012] Optionally, the second initial single-layer active sperm image and the structure information image are substituted into a single residual block to obtain a basic feature and a structure basic feature, respectively, including: The second initial single-layer active sperm image and the structure information image are substituted into a single residual block to obtain a first feature and a second feature, respectively; The first feature and the second feature are subjected to feature compression through average pooling and maximum pooling, and are merged to obtain a fusion feature; The fusion feature, the first feature and the second feature are spliced to obtain a joint feature; The joint feature is subjected to a 5x5 convolution layer plus Sigmoid activation to generate a first weight and a second weight; the first weight is multiplied by the first feature to obtain a basic feature; and the second weight is multiplied by the second feature to obtain a structure basic feature.

[0013] Optionally, the basic feature and the structure basic feature are substituted into a multi-scale sampling model to obtain a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature, including: The basic feature and the structure basic feature are fused to obtain a fusion basic feature; The fusion basic feature is split into a first sub-fusion basic feature and a second sub-fusion basic feature through a 1x1 convolution; The first sub-fusion basic feature is subjected to convolution operation through a first convolution kernel to obtain a first sub-convolution feature; and the second sub-fusion basic feature is subjected to convolution operation through a second convolution kernel to obtain a second sub-convolution feature; The second sub-convolution feature is subjected to 3x3 convolution plus 1x1 convolution plus Sigmoid activation to obtain a second sub-convolution attention; A second sub-attention feature is obtained according to the second sub-convolution attention multiplied by the second sub-convolution feature; The second sub-attention feature and the first sub-convolution feature are spliced and then subjected to 3x3 convolution and residual connection to obtain an enhanced multi-stage feature; the multi-stage feature includes a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature.

[0014] Optionally, the low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature are substituted into the double-path feature optimization model to obtain a low-scale final fusion feature, a medium-scale final fusion feature and a high-scale final fusion feature, including: The first scale feature and the second scale feature are subjected to 1*1 convolution, then the second scale feature is subjected to up-sampling, and the first scale feature is subjected to feature fusion and bilinear interpolation to generate a first semantic feature; the first scale feature is any one of the medium-scale morphological feature and the high-scale morphological feature; the second scale feature is any one of the medium-scale morphological feature and the low-scale morphological feature; the scale of the second scale feature is smaller than that of the first scale feature; The first semantic feature is subjected to 1*1 convolution and substituted into a Sigmoid activation function to obtain a foreground attention map and a background attention map; The foreground attention map is multiplied by the second scale feature to obtain a foreground path, and the background attention map is multiplied by the second scale feature to obtain a background path; when the first scale feature is the high-scale morphological feature and the second scale feature is the medium-scale morphological feature, a first foreground path and a first background path are obtained; when the first scale feature is the high-scale morphological feature and the second scale feature is the low-scale morphological feature, a second foreground path and a second background path are obtained; when the first scale feature is the medium-scale morphological feature and the second scale feature is the low-scale morphological feature, a third foreground path and a third background path are obtained; The first foreground path, the second foreground path and the third foreground path are fused to obtain a fused foreground path; The first background path, the second background path and the third background path are fused to obtain a fused background path; The fused foreground path and the fused background path are respectively optimized through a residual block, spliced and subjected to 1*1 convolution to generate a final fusion feature; The low-scale morphological feature and the final fusion feature are averaged to obtain a low-scale final fusion feature; the medium-scale morphological feature and the final fusion feature are averaged to obtain a medium-scale final fusion feature; and the high-scale morphological feature and the final fusion feature are averaged to obtain a high-scale final fusion feature.

[0015] Optionally, the first magnification operation is implemented by using a microscope; the microscope uses a 60x objective lens and a 10x ocular lens.

[0016] In the second aspect of the implementation of the present application, a high-precision sperm morphological screening system is provided, comprising: A filtering module is configured to obtain raw semen and filter the raw semen to obtain initial semen; An initial single-layer active sperm screening module is configured to spread the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm; The first amplification module is configured to perform a first amplification operation on the initial single-layer active sperm to obtain a first initial single-layer active sperm image. The amplified initial single-layer active sperm image generation module is configured to amplify the first initial single-layer active sperm image to a target resolution to obtain an amplified initial single-layer active sperm image. The second amplification module is configured to perform a second amplification operation on the first initial single-layer active sperm image according to the amplified initial single-layer active sperm image to obtain a second initial single-layer active sperm image. The structural information image generation module is configured to obtain a structural information image by using a Sobel operator for the second initial single-layer active sperm image. The high-precision sperm morphology selection module is configured to input the second initial single-layer active sperm image and the structural information image into a high-precision sperm morphology selection model to obtain target high-quality sperm.

[0017] The present application has the following advantages: The present application provides a high-precision sperm morphology screening method, which can effectively remove impurities, dead sperm and other useless components in the original semen by filtering the original semen to obtain initial semen, then spreading the initial semen into a preset single-layer sperm liquid strip to ensure uniform distribution of sperm to form initial single-layer active sperm, avoiding the influence of sperm stacking on observation and screening. Then, through two amplification operations, a first initial single-layer active sperm image is obtained, and then amplified to a target resolution, and based on this, a second initial single-layer active sperm image is optimized, gradually improving the image clarity and detail presentation, making it easier to capture sperm morphology characteristics. Then, with the help of a Sobel operator, a structural information image is extracted, which can further highlight the key structural features of the sperm, such as outline and texture, providing more abundant judgment basis for accurate identification. Finally, the second initial single-layer active sperm image and the structural information image are input into a high-precision sperm morphology selection model to accurately screen out target high-quality sperm with excellent morphology and strong vitality, greatly improving the screening efficiency and accuracy of high-quality sperm, not just simply identifying headless / less vacuole sperm, providing a reliable source of high-quality sperm for related fields such as assisted reproduction, ensuring the success rate of subsequent cultivation and other links, and improving the screening efficiency of high-precision sperm. BRIEF DESCRIPTION OF DRAWINGS

[0018] The present application will be further described below with reference to the accompanying drawings.

[0019] Figure 1 A flowchart of a high-precision sperm morphology screening method provided by the present application embodiment; Figure 2 A flowchart of a high-precision sperm morphology screening method provided by the present application embodiment; Figure 3A sperm image under different magnifications is provided for the embodiment of the present application. Figure 4 A deformed sperm contrast image is provided for the embodiment of the present application. Figure 5 A high-precision sperm morphology screening system framework is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application.

[0021] Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] The embodiment of the present application provides a high-precision sperm morphology screening method. Referring to Figure 1 , Figure 1 A flowchart of a high-precision sperm morphology screening method is provided for the embodiment of the present application. The method comprises the following steps: S101, obtaining raw semen and filtering the raw semen to obtain initial semen; S102, spreading the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm; S103, performing a first magnification operation on the initial single-layer active sperm to obtain a first initial single-layer active sperm image; S104, magnifying the first initial single-layer active sperm image to a target resolution to obtain a magnified initial single-layer active sperm image; S105, performing a second magnification operation on the first initial single-layer active sperm image according to the magnified initial single-layer active sperm image to obtain a second initial single-layer active sperm image; S106, obtaining a structure information image through a Sobel operator for the second initial single-layer active sperm image; S107, substituting the second initial single-layer active sperm image and the structure information image into a high-precision sperm morphology selection model to obtain a target high-quality sperm.

[0023] Based on the high-precision sperm morphology screening method provided by this invention, the initial semen is first filtered to remove impurities, dead sperm, and other useless components. Then, it is spread into a pre-set monolayer sperm strip to ensure uniform sperm distribution, forming an initial monolayer of motile sperm and preventing stacking that could affect observation and screening. Subsequently, two magnification operations are performed: first, a first image of the initial monolayer of motile sperm is obtained; then, it is magnified to the target resolution and optimized to obtain a second image, gradually improving clarity and detail to facilitate the capture of sperm morphological features. Next, the Sobel operator is used to extract structural information images, highlighting key features such as sperm outline and texture, providing a basis for accurate identification. Finally, the two images are substituted into a high-precision sperm morphology selection model to accurately screen out high-quality sperm with excellent morphology and strong motility. This method not only simply determines the absence or scarcity of vacuoles in the sperm head but also significantly improves the efficiency and accuracy of high-quality sperm screening, providing a reliable source of high-quality sperm for assisted reproduction and other fields, ensuring a high success rate for subsequent culture.

[0024] In one implementation, the original semen is from patients with severe teratospermia or occult azoospermia; the patient's sperm is liquefied at 37°C for 30-60 minutes, 20 microliters of liquefied semen are collected for microscopic examination, and the sperm motility and quantity are recorded (the quantity is observed and counted using existing microscopes).

[0025] In one implementation, if the original semen sample comes from a patient with severe teratospermia, the sperm density is determined to be below 1×10⁻⁶. 6 (Unit: quantity / ml) If the sperm density is less than 1×10 6 If the sperm density is equal to or higher than 1×10⁻⁶, then the initial semen is obtained by direct washing; 6 Then, the semen is subjected to density gradient centrifugation (300g / 20min) and the precipitate is retained to obtain the initial semen.

[0026] In one implementation, see [link to implementation details]. Figure 2 , Figure 2 A flowchart illustrating a high-precision sperm morphology screening method is provided. Referring to Part I (implementation step S101), semen is first collected. Gradient solution 1 (90% colloidal silica particles) and gradient solution 2 (45% colloidal silica particles) are added to the semen, followed by centrifugation (centrifugation time is 20 min). The supernatant is removed from the centrifuged mixture to obtain sperm precipitate. Culture medium (G-IVF plus) is added to the sperm precipitate, and the mixture is mixed and washed to obtain initial semen. After obtaining the initial semen, referring to Part II (implementation step S102), the initial semen is placed in a glass dish and filtered using a monolayer sperm strip method to obtain an initial monolayer of motile sperm. After obtaining the initial monolayer of motile sperm, referring to Part III (implementation steps S103-S105).

[0027] In one implementation, the original semen is the semen of an azoospermia patient, and after being liquefied at 37°C for 30-60 minutes, the supernatant is removed and 100 microliters of the precipitate is obtained as the initial semen.

[0028] In one implementation, the first magnification operation is specifically that, after the initial single-layer motile sperm obtained by adding to the PVP droplet, the initial single-layer motile sperm is magnified and digital images are collected by using the objective lens (60 times) and eyepiece (10 times) dedicated to the IMSI system (for example, refer to the cell screening manual published by the German enterprise OCTAX Microscience GmbH in 2008) to obtain the first initial single-layer motile sperm image; the first magnification operation is implemented by using a microscope; the microscope uses a 60 times objective lens and a 10 times eyepiece; at this time, the first initial single-layer motile sperm image can be magnified 600 times by combining the objective lens and the eyepiece; and then the second initial single-layer motile sperm image is obtained by performing a second magnification operation on the first initial single-layer motile sperm image (600 times); the second magnification operation is realized by software to magnify the first initial single-layer motile sperm image (600 times) to the second initial single-layer motile sperm image (2000-6000 times).

[0029] In one implementation, referring to Figure 3 , Figure 3 A sperm image under different magnifications is provided, and the magnification under ICSI is 200 times and 300 times, at which time only the shape of the sperm can be seen, and the internal information of the sperm cannot be observed; the sperm image of 1200 times collected by SLS-IMSI (single-layer sperm strip method combined with ultra-high resolution preferred microinjection technology) can roughly see the morphological structure of the sperm, such as the fine features of the sperm head and tail, which is very crucial for evaluating the quality of the sperm and screening the morphologically normal sperm.

[0030] In one implementation, referring to Figure 4 , Figure 4 A contrast image of abnormal sperm is provided, in which the abnormal sperm is indicated, and the normal sperm is on the right side of the image.

[0031] In one implementation, the target high-quality sperm is the sperm with symmetrical head morphology, oval structure of the head, homogeneity of nuclear chromatin not more than 1 vacuole, or the area of the vacuole is less than 4% of the area of the nucleus, and normal appearance of the nucleus, and the limits of the average length and width of the sperm are length: 4.75±0.28 μm, and width: 3.28±0.20 μm.

[0032] In one implementation, the original semen is filtered to remove basic contaminants such as seminal plasma impurities, dead sperm, and abnormal fragments, to obtain initial semen, thereby reducing interference for subsequent screening; and a preset single-layer sperm strip laying method is used for filtering, in which a standardized strip laying method is used to further separate the initial single-layer motile sperm, and this step can selectively retain sperm with strong motility and eliminate immotile sperm and weak sperm, thereby completing the first round of quality screening from the dimension of motility and laying a foundation for subsequent morphological analysis.

[0033] In one implementation, two magnification operations are performed, the first magnification is performed by an objective lens and an ocular lens, and the second magnification is performed by a software AI model, thereby gradually focusing on the individual sperm from a macroscopic view to a microscopic view, ensuring that the subsequent morphological analysis can clearly capture the details of the head, neck, and tail of the sperm, and avoiding misjudgment caused by insufficient image resolution.

[0034] In one implementation, compared with manual morphological evaluation, the high-precision sperm morphological selection model can establish a standardized judgment standard based on a large amount of labeled data, accurately identify high-quality sperm with a regular head, no deformity, and a complete tail, reduce human error, cover subtle morphological features that are difficult for humans to detect, and finally ensure that the target high-quality sperm selected has better morphological integrity.

[0035] In one embodiment, the initial semen is filtered by the preset single-layer sperm strip laying method to obtain initial single-layer motile sperm, including: Step 1, glass dish preparation: preparing N Gamete liquid strips of 20 uL and one drop of 2 uL multi-antennae PVP liquid drop in an IMSI glass dish; Step 2, semen strip laying: for each Gamete liquid strip, 10 uL of liquid in the Gamete liquid strip is sucked, and 10 uL of initial semen is injected, and centrifugal semen strip is obtained by centrifugation for a preset period of time; Step 3, preliminary screening of sperm: under an inverted microscope, all initial normal sperm in the centrifugal semen strip are selected by an ICSI needle according to a preset rule, and are added to the PVP liquid drop to obtain initial single-layer motile sperm.

[0036] In one implementation, the glass dish preparation of the improved single-layer sperm strip method uses IMSI (version number octaxeyeware) special glass dish (OOPW-IC03, OosafeICSI, IMSIDish, etc.), and the IMSI system needs to use a cell dish made of glass material with better light transmission. The glass injection dish is prepared at least 2 hours before the formal IMSI. In the special dish, use a micropipette gun to prepare three Gamete liquid strips with a volume of 20 microliters, a length of 4 cm, and a width of 0.5 cm by connecting the head and tail. Two 10-microliter Gamete droplets are placed on the lower side of the dish for placing eggs, and a 2-microliter multi-antennae PVP droplet is placed in the center of the dish for placing sperm. Finally, cover it with paraffin oil and place it in the incubator for balancing.

[0037] In one implementation, the single-layer sperm strip method collects a small amount of sperm; first, use a micropipette gun to suck 10 microliters of Gamete from the liquid strip, then suck 10 microliters of semen and centrifuge the sediment, and slowly add the sediment from the head to the tail. Each liquid is slowly filled with semen sediment according to the above method. After all the semen sediment is completely laid, place it in the incubator for 15 minutes, then take out the glass dish and observe whether there are active sperm swimming out to the edges of the liquid strip.

[0038] In one implementation, the single IMSI operation sperm preliminary screening; after the strip is completely laid, culture for 15-30 min, collect all the morphologically normal sperm into the PVP droplet under the inverted microscope. If the number of sperm swimming out is particularly small, you can go to the middle of the liquid strip to find whether there are micro-vibration active sperm as the initial single-layer active sperm; the preset time period is 15-30 min; the preset rule is that the sperm head is symmetrical and the head is elliptical structure.

[0039] In one embodiment, the second magnification operation on the first initial single-layer active sperm image according to the magnified initial single-layer active sperm image to obtain a second initial single-layer active sperm image comprises: Substitute the first initial single-layer active sperm image and the magnified initial single-layer active sperm image into the pre-trained VAE encoder to obtain a low-resolution latent vector and a high-resolution latent vector, respectively; Add noise to the low-resolution latent vector and the high-resolution latent vector to obtain a low-resolution noise latent vector and a high-resolution noise latent vector, respectively; Update the high-resolution noise latent vector according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector; Generate a mean square error loss according to the low-resolution latent vector and the updated high-resolution latent vector; The initial single-layer motile sperm image is updated according to the mean square error loss to obtain a second initial single-layer motile sperm image.

[0040] In an implementation, the target resolution is a resolution that needs to be magnified by a technician, and the first initial single-layer motile sperm image is usually magnified by 10 times; the pre-trained VAE encoder is a variational autoencoder; and the low-resolution latent vector and the high-resolution latent vector are added with noise, which is standard normal distribution noise.

[0041] In an implementation, the mean square error loss is generated according to the low-resolution latent vector and the updated high-resolution latent vector, specifically, wavelet transform is performed on the low-resolution latent vector and the updated high-resolution latent vector respectively, and low-frequency parts corresponding to the low-resolution latent vector and the updated high-resolution latent vector are extracted to obtain low-frequency low-resolution latent features and low-frequency high-resolution latent features, and then the square of the difference between the low-frequency low-resolution latent features and the low-frequency high-resolution latent features is calculated to obtain the mean square error loss.

[0042] In an implementation, the initial single-layer motile sperm image is updated according to the mean square error loss to obtain a second initial single-layer motile sperm image, specifically, the second initial single-layer motile sperm image is obtained by the formula , wherein A is the high-resolution latent vector corresponding to the second initial single-layer motile sperm image, A0 is the high-resolution latent vector of the magnified initial single-layer motile sperm image, is a cosine decay schedule, the high-resolution latent vector of the magnified initial single-layer motile sperm image is gradiented, is the mean square error loss; after the high-resolution latent vector corresponding to the second initial single-layer motile sperm image is calculated, the second initial single-layer motile sperm image is obtained by substituting the decoder of the variational autoencoder.

[0043] In an implementation, the mean square error loss is calculated by the low-resolution latent vector and the updated high-resolution latent vector in the process, and the essence is to constrain the generation of the high-resolution image by using the global structure of the low-resolution image. This constraint can effectively avoid the abnormal head shape in the traditional magnification method, ensure that the magnified sperm image is consistent with the original low-resolution image in the global shape, and meet the core demand of the biomedical image for structural authenticity; the head of the sperm in the low-resolution image is elliptical and smoothly connected with the tail, and through the loss constraint, the high-resolution image will retain this global feature and will not be damaged by the detail generation.

[0044] In an implementation, the operation of adding noise to the low-resolution and high-resolution latent vectors draws on the noise and denoising logic of the diffusion model: the noise provides flexibility for generating details of the high-resolution image, and subsequent updating of the high-resolution noise latent vector by the low-resolution noise latent vector; for sperm images, this does not only generate blurred fine structures in the low-resolution image, but also does not generate details that do not conform to biological laws, and both magnification clarity and biological rationality are taken into account.

[0045] In an implementation, the iterative optimization of the high-resolution image is achieved through a closed loop of mean square error loss and gradient update, each round of update makes the high-resolution image closer to the global structure of the low-resolution image, while retaining reasonable details, and the final output image achieves a better balance between structural consistency and detail richness. This optimization can make the magnified image clearly present the tail swing amplitude, without distorting the head shape, and improve the accuracy of subsequent analysis.

[0046] In an embodiment, updating the high-resolution noise latent vector according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector comprises: The first initial single-layer active sperm image is segmented into a set of labeled sub-initial single-layer active sperm images according to a preset first region size, and a text prompt is generated for each sub-initial single-layer active sperm image by a large language model; The low-resolution latent noise vector and the text prompt corresponding to each sub-initial single-layer active sperm image are substituted into a preset Unet network to obtain a low-resolution attention score corresponding to each sub-initial single-layer active sperm image; The high-resolution latent vector is segmented into a set of labeled sub-high-resolution latent vectors according to a preset second region size; the labels in the set of sub-high-resolution latent vectors correspond one-to-one to the labels in the set of sub-initial single-layer active sperm images; The text prompt, the low-resolution attention score and the sub-high-resolution latent vector corresponding to the same label are substituted into an attention synthesizer for denoising to obtain a sub-updated high-resolution latent vector corresponding to the label; The sub-updated high-resolution latent vectors corresponding to all labels are merged to obtain an updated high-resolution latent vector.

[0047] In an implementation, the first initial single-layer active sperm image is divided into sub-regions with a preset first region size, to obtain a set of labeled sub-initial single-layer active sperm images and a set of labeled sub-high-resolution latent vectors by dividing the high-resolution latent vector into sub-regions with a preset second region size; the set of labeled sub-initial single-layer active sperm images and the set of labeled sub-high-resolution latent vectors correspond to each other in position; first, the low-resolution first initial single-layer active sperm image is divided into sub-images according to the preset first region size, and the label of each sub-image is bound to the spatial position (such as the row and column) of the sub-image in the low-resolution image (for example, the first sub-image in the upper left corner is labeled as 1, the adjacent block on the right side is labeled as 2, and so on); when the high-resolution latent vector is divided into sub-regions according to the preset second region size, the corresponding position of the low-resolution sub-image in the high-resolution latent space is determined through the equal proportion scaling from low resolution to high resolution and the VAE compression ratio mapping from high-resolution pixel space to latent space, and the sub-high-resolution latent vector in the position is labeled with the same label, to finally realize the sub-regions with the same label.

[0048] In an implementation, the preset Unet network is a Unet with an attention synthesizer; different regions (such as the head, tail, and connecting part) of the sperm image have completely different biological characteristics (the head needs to retain the elliptical acrosome, and the tail needs to present a slender and unbroken texture). The image is divided into sub-regions with labels by dividing the image into sub-regions with a preset region size, and a dedicated text prompt is generated for each sub-region (such as the head acrosome region, the first 1 / 3 is arc-shaped, the middle section of the tail, and the fiber texture is clear), which can enable the attention mechanism to accurately focus on the region-specific characteristics; after the low-resolution attention score is aligned with the high-resolution sub-region by the label, the tail region can avoid learning the acrosome characteristics of the head, ensuring that the denoising process of each region is only guided by its own exclusive characteristics, and preventing cross-region detail confusion.

[0049] In an implementation, the attention synthesizer simultaneously receives a text prompt (semantic constraint), a low-resolution attention score (low-resolution feature guidance), and a sub-high-resolution latent vector (to-be-optimized high-resolution feature), to realize multi-source information fusion denoising: the text prompt provides what should be generated in this region (such as no broken tail), to avoid biological errors; the low-resolution attention score provides which features should be focused on in this region (such as high weight for the elliptical feature in the head region), to ensure consistency with the low-resolution image; and the sub-high-resolution latent vector provides the to-be-optimized basic feature, to gradually generate high-resolution details through denoising; this fusion mechanism enables the local details of the sperm (such as the acrosome edge definition and the tail fiber texture) to conform to biological laws and be faithful to the feature distribution of the original low-resolution image.

[0050] In an embodiment, the second initial single-layer active sperm image and the structure information image are substituted into the high-precision sperm morphology selection model to obtain a target high-quality sperm. The second initial single-layer active sperm image and the structure information image are substituted into the single residual block respectively to obtain a basic feature and a structure basic feature; The basic feature and the structure basic feature are substituted into the multi-scale sampling model to obtain a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature; The low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature are substituted into the double-path feature optimization model to obtain a low-scale final fusion feature, a medium-scale final fusion feature and a high-scale final fusion feature; According to the low-scale final fusion feature, the medium-scale final fusion feature and the high-scale final fusion feature, a target high-quality sperm is screened in combination with a morphological quality standard.

[0051] In an implementation, the structure information image extracted by the Sobel operator is complementary to the original sperm image: the original image is good at retaining biological characteristics such as sperm head staining depth, and the structure information image strengthens the morphological edge; after the single residual block processing, the basic feature and the structure basic feature respectively retain the integrity of the two types of information, laying a feature non-missing foundation for subsequent multi-scale analysis.

[0052] In an implementation, the second initial single-layer active sperm image and the structure information image are substituted into the single residual block respectively to obtain a first feature and a second feature; the first feature and the second feature are subjected to feature compression through average pooling and maximum pooling, and are merged to obtain a fusion feature; the fusion feature, the first feature and the second feature are spliced to obtain a joint feature, and the joint feature is subjected to 5x5 convolution layer plus Sigmoid activation to generate a first weight and a second weight; the first weight is multiplied by the first feature to obtain the basic feature; and the second weight is multiplied by the second feature to obtain the structure basic feature.

[0053] In an implementation, the basic visual features (such as sperm morphology and contour) of the second initial single-layer active sperm image and the detailed features (such as internal texture) of the structure information image are extracted respectively, and then the complete fusion of the two types of information is realized through splicing to avoid the loss of key details of a single type of feature. The first and second weights generated through 5x5 convolution plus Sigmoid can automatically judge the importance of the two types of features in the current task (such as sperm motility evaluation), for example, when the structure details are more critical for judgment, the second weight will increase, allowing the structure basic feature to contribute more and reducing the interference of useless features.

[0054] In an implementation, the operation of substituting the basic feature and the structural basic feature into the multi-scale sampling model to obtain the low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature is specifically: fusing the basic feature and the structural basic feature to obtain a fused basic feature; splitting the fused basic feature into two branches through 1x1 convolution to obtain a first sub-fused basic feature and a second sub-fused basic feature; performing convolution operation on the first sub-fused basic feature with a first convolution kernel (the convolution kernel is 3) to obtain a first sub-convolution feature; performing convolution operation on the second sub-fused basic feature with a second convolution kernel (the convolution kernel is 5) to obtain a second sub-convolution feature; performing 3x3 convolution plus 1x1 convolution plus Sigmoid activation on the second sub-convolution feature to obtain a second sub-convolution attention, and multiplying the second sub-convolution attention by the second sub-convolution feature to obtain a second sub-attention feature; performing 3x3 convolution plus residual connection (added to the fused basic feature) on the second sub-attention feature and the first sub-convolution feature after splicing to generate an enhanced multi-stage feature (the low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature).

[0055] In an implementation, the division of the low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature accurately matches the scale difference of different parts of the sperm; the low-scale feature captures the fine structure of the tail, and whether there are multiple deformities (such as short tail, curled tail, folded tail, no tail, irregular tail); the medium-scale feature captures the neck information and identifies whether the arrangement of the mitochondria in the neck is abnormal; the high-scale feature focuses on the size of the head, determines whether there is an acrosome in the head, whether the acrosome morphology is regular, and whether there are vacuoles in the head; the combination of the three achieves complete feature coverage of the global morphology of the sperm, and solves the problem of traditional single-scale analysis.

[0056] In an implementation, after splicing different branch features, 3x3 convolution is further used for fusion and combined with residual connection (added to the original fused basic feature), which not only strengthens the relevance of the head, neck and tail of the sperm in the multi-scale feature, but also avoids the feature degradation problem in the deep network.

[0057] In one implementation, the low-scale morphological features, the medium-scale morphological features and the high-scale morphological features are substituted into the double-path feature optimization model to obtain low-scale final fusion features, medium-scale final fusion features and high-scale final fusion features. Specifically, the first scale features and the second scale features are first subjected to 1x1 convolution, and then the second scale features are up-sampled, and the first scale features and the second scale features are subjected to feature fusion and bilinear interpolation to generate first semantic features; the first semantic features are subjected to 1x1 convolution and substituted into a Sigmoid activation function to obtain a foreground attention map A, and thus 1-A obtains a background attention map; the foreground attention map is multiplied by the second scale features to obtain a foreground path, and the background attention map is multiplied by the second scale features to obtain a background path; when the first scale features are high-scale morphological features and the second scale features are medium-scale morphological features, first foreground path and first background path are obtained; when the first scale features are high-scale morphological features and the second scale features are low-scale morphological features, second foreground path and second background path are obtained; when the first scale features are medium-scale morphological features and the second scale features are low-scale morphological features, third foreground path and third background path are obtained; the first foreground path, the second foreground path and the third foreground path are fused to obtain a fused foreground path; the first background path, the second background path and the third background path are fused to obtain a fused background path; the fused foreground path and the fused background path are respectively optimized through a residual block, and after splicing, 1x1 convolution is performed to generate final fusion features, and the low-scale morphological features and the final fusion features are averaged to obtain low-scale final fusion features; the medium-scale morphological features and the final fusion features are averaged to obtain medium-scale final fusion features; and the high-scale morphological features and the final fusion features are averaged to obtain high-scale final fusion features.

[0058] In one implementation, through the interaction of high-medium, high-low and medium-low scales (first convolution and alignment of channels, and then up-sampling to match the size), different scale features are deeply fused in the semantic level. The global structure information of the high scale can guide the optimization of the local features of the medium and low scales, and avoid one-sidedness of single scale features.

[0059] In one implementation, a foreground-background attention mechanism is introduced (a foreground attention map A and a background attention map B are generated through Sigmoid), and the foreground attention map A and the background attention map B are multiplied by the second scale features to obtain a foreground path and a background path, respectively. Figure 1A), which can automatically distinguish sperm targets (foreground) from background noise; the foreground path enhances sperm-related features (such as tail swing area and head shape), and suppresses irrelevant background interference; the background path can assist model learning, and indirectly improve the purity of foreground features; this separation mechanism is particularly important for small targets such as sperm, which are easily disturbed by the background. In one implementation, according to the low-scale final fusion feature, the medium-scale final fusion feature, and the high-scale final fusion feature, a target high-quality sperm is screened by combining a morphological quality standard, specifically, a "anchor-based detection head" is deployed on the low-scale final fusion feature, the medium-scale final fusion feature, and the high-scale final fusion feature, and the detection result is output by classification (whether it is a high-quality sperm) and bounding box regression (locating the position of the sperm); the high-quality morphological quality standard screening: the sperm head shape is symmetrical, the head is an oval structure, the nuclear chromatin has homogeneity of not more than 1 vacuole, or the vacuole area is less than 4% of the nuclear area, the normal nuclear appearance, the average length and width of the sperm are limited to length: 4.75±0.28μm, width: 3.28±0.20μm, and the sperm that meets the above conditions is marked as a target high-quality sperm.

[0060] In one implementation, the dual-path feature optimization model uses parallel processing of foreground enhancement and background suppression: the foreground path focuses on the core area of the sperm, and amplifies the weight of the effective morphological features; the background path filters interference information such as cell debris and bubbles, to avoid noise being misjudged as sperm features; the final output fusion feature retains key morphological details and reduces redundant information interference, making subsequent screening more accurate.

[0061] Based on the same inventive concept, the embodiments of the present application also provide a high-precision sperm morphology screening system. Referring to Figure 5 , Figure 5 A framework diagram of a high-precision sperm morphology screening system provided by the embodiments of the present application, comprising: A filtering module for obtaining raw semen, filtering the raw semen to obtain initial semen; An initial single-layer motile sperm screening module for spreading the initial semen into a pre-set single-layer sperm liquid strip to obtain initial single-layer motile sperm; A first magnification module for performing a first magnification operation on the initial single-layer motile sperm to obtain a first initial single-layer motile sperm image; A magnified initial single-layer motile sperm image generation module for magnifying the first initial single-layer motile sperm image to a target resolution to obtain a magnified initial single-layer motile sperm image; A second magnification module for performing a second magnification operation on the first initial single-layer motile sperm image according to the magnified initial single-layer motile sperm image to obtain a second initial single-layer motile sperm image; a structure information image generation module configured to obtain a structure information image from the second initial single-layer active sperm image by using a Sobel operator; a high-precision sperm morphology selection module configured to obtain a target high-quality sperm by inputting the second initial single-layer active sperm image and the structure information image into a high-precision sperm morphology selection model.

[0062] Based on the high-precision sperm morphology screening system provided in the embodiments of the present application, the original semen is first filtered to obtain initial semen, and impurities, dead sperm and other useless components are removed; then the initial semen is spread into a preset single-layer sperm liquid strip, so that the sperm is uniformly distributed to form an initial single-layer active sperm, and the stacking is avoided to affect the observation and screening. Subsequently, two magnification operations are performed: first, an initial single-layer active sperm image is obtained, and then the image is magnified to a target resolution and optimized to obtain a second image, so that the clarity and details are gradually improved, and the sperm morphology features are easily captured. Then, a structure information image is extracted by using a Sobel operator, and key features such as the outline and texture of the sperm are highlighted, thereby providing a basis for accurate identification. Finally, the two images are input into a high-precision sperm morphology selection model, and a target high-quality sperm with excellent morphology and strong motility is accurately screened out. Not only the head without / few vacuoles is simply judged, but also the screening efficiency and accuracy of high-quality sperm are greatly improved, thereby providing a reliable source of high-quality sperm for the field of assisted reproduction and the like, and ensuring the success rate of subsequent cultivation.

[0063] In one embodiment, the initial single-layer active sperm screening module includes: a glass dish preparation module configured to prepare N Gamete liquid strips of 20 uL and a PVP droplet of 2 uL in an IMSI glass dish; a semen strip spreading module configured to, for each Gamete liquid strip, suck 10 uL of liquid from the Gamete liquid strip, inject 10 uL of initial semen, and obtain a centrifuged semen strip by centrifugal sedimentation for a preset period of time; an initial sperm screening module configured to pick up all initial normal sperm in the centrifuged semen strip by using an ICSI needle under an inverted microscope according to a preset rule, and add the sperm to the PVP droplet to obtain an initial single-layer active sperm.

[0064] In one embodiment, the second magnification module includes: a magnified initial single-layer active sperm image determination module configured to magnify the first initial single-layer active sperm image to a target resolution to obtain a magnified initial single-layer active sperm image; an image encoding module configured to input the first initial single-layer active sperm image and the magnified initial single-layer active sperm image into a pre-trained VAE encoder to obtain a low-resolution latent vector and a high-resolution latent vector, respectively; a noise adding module, configured to add noise to the low-resolution latent vector and the high-resolution latent vector respectively to obtain a low-resolution noise latent vector and a high-resolution noise latent vector; a high-resolution noise latent vector updating module, configured to update the high-resolution noise latent vector according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector; a mean square error loss calculation module, configured to generate a mean square error loss according to the low-resolution latent vector and the updated high-resolution latent vector; a gradient updating module, configured to perform gradient updating on the amplified initial single-layer motile sperm image according to the mean square error loss to obtain a second initial single-layer motile sperm image.

[0065] In an embodiment, the high-resolution noise latent vector updating module comprises: a text prompt generation module, configured to segment the first initial single-layer motile sperm image by a preset first region size to obtain a set of sub-initial single-layer motile sperm images with labels, and generate a corresponding text prompt for each sub-initial single-layer motile sperm image by a large language model; an attention score generation module, configured to input the low-resolution noise latent vector and the text prompt corresponding to each sub-initial single-layer motile sperm image into a preset Unet network to obtain a low-resolution attention score corresponding to each sub-initial single-layer motile sperm image; a high-resolution latent vector segmentation module, configured to segment the high-resolution latent vector by a preset second region size to obtain a set of sub-high-resolution latent vectors with labels; the labels in the set of sub-high-resolution latent vectors correspond one-to-one to the labels in the set of sub-initial single-layer motile sperm images; a sub-updated high-resolution latent vector denoising module, configured to input the text prompt, the low-resolution attention score and the sub-high-resolution latent vector corresponding to the same label into an attention synthesizer for denoising to obtain a sub-updated high-resolution latent vector corresponding to the label; a sub-updated high-resolution latent vector merging module, configured to merge the sub-updated high-resolution latent vectors corresponding to all labels to obtain the updated high-resolution latent vector.

[0066] In an embodiment, the high-precision sperm morphology selection module comprises: a basic feature extraction module, configured to input the second initial single-layer motile sperm image and the structural information image into a single residual block respectively to obtain a basic feature and a structural basic feature; a multi-scale feature extraction module, configured to input the basic feature and the structural basic feature into a multi-scale sampling model to obtain a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature; The multi-scale feature fusion module is configured to input the low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature into a double-path feature optimization model to obtain a low-scale final fusion feature, a medium-scale final fusion feature and a high-scale final fusion feature. The target high-quality sperm screening module is configured to screen target high-quality sperm according to the low-scale final fusion feature, the medium-scale final fusion feature and the high-scale final fusion feature in combination with a morphological quality standard.

[0067] The above describes one embodiment of the present application in detail, but the content is only a preferred embodiment of the present application and cannot be considered as limiting the implementation range of the present application. Any equivalent changes and improvements made according to the application scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A high precision method for sperm morphology screening, characterized by, The method comprises: obtaining raw semen, filtering the raw semen to obtain initial semen; spreading the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm; performing a first magnification operation on the initial single-layer active sperm to obtain a first initial single-layer active sperm image; magnifying the first initial single-layer active sperm image to a target resolution to obtain a magnified initial single-layer active sperm image; performing a second magnification operation on the first initial single-layer active sperm image according to the magnified initial single-layer active sperm image to obtain a second initial single-layer active sperm image; obtaining a structure information image through a Sobel operator for the second initial single-layer active sperm image; inputting the second initial single-layer active sperm image and the structure information image into a high-precision sperm morphology selection model to obtain target high-quality sperm.

2. The method of claim 1, wherein the method is a high precision method of sperm morphology screening. Spreading the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm comprises: Step 1, glass dish preparation: preparing N 20uL Gamete liquid strips and a drop of 2uL multi-edge antenna-like PVP droplet in IMSI glass dishes; Step 2, semen spreading: for each Gamete liquid strip, 10uL of liquid is sucked out of the Gamete liquid strip, and 10uL of initial semen is injected, and centrifugal semen spreading is performed through a preset time period to obtain centrifugal semen spreading; Step 3, preliminary screening of sperm: under an inverted microscope, all initial normal sperm in the centrifugal semen spreading are selected by a preset rule using an ICSI needle, and are added to the PVP droplet to obtain initial single-layer active sperm.

3. The method of claim 1, wherein the method is a high precision method of sperm morphology screening. Performing a second magnification operation on the first initial single-layer active sperm image according to the magnified initial single-layer active sperm image to obtain a second initial single-layer active sperm image comprises: respectively inputting the first initial single-layer active sperm image and the magnified initial single-layer active sperm image into a pre-trained VAE encoder to obtain a low-resolution latent vector and a high-resolution latent vector; respectively adding noise to the low-resolution latent vector and the high-resolution latent vector to obtain a low-resolution noise latent vector and a high-resolution noise latent vector; updating the high-resolution noise latent vector according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector; generating a mean square error loss according to the low-resolution latent vector and the updated high-resolution latent vector; performing gradient update on the magnified initial single-layer active sperm image according to the mean square error loss to obtain a second initial single-layer active sperm image.

4. The high precision sperm morphology screening method according to claim 3, wherein, Updating the high-resolution noise latent vector according to the low-resolution noise latent vector to obtain an updated high-resolution latent vector comprises: segmenting the first initial single-layer active sperm image by a preset first area size to obtain a set of labeled sub-initial single-layer active sperm images, and generating a corresponding text prompt for each sub-initial single-layer active sperm image through a large language model; inputting the low-resolution noise latent vector and the text prompt corresponding to each sub-initial single-layer active sperm image into a preset Unet network to obtain a low-resolution attention score corresponding to each sub-initial single-layer active sperm image; The high-resolution latent vector is segmented by a preset second region size to obtain a labeled sub-high-resolution latent vector set; the label in the sub-high-resolution latent vector set corresponds to the label in the sub-initial single-layer active sperm image set one by one; The text prompt, the low-resolution attention score and the sub-high-resolution latent vector corresponding to the same label are substituted into the attention synthesizer to obtain a sub-updated high-resolution latent vector corresponding to the label; The sub-updated high-resolution latent vectors corresponding to all labels are merged to obtain an updated high-resolution latent vector.

5. The method of claim 1, wherein the method is a high precision method of sperm morphology screening. The second initial single-layer active sperm image and the structural information image are substituted into a high-precision sperm morphology selection model to obtain a target high-quality sperm, including: The second initial single-layer active sperm image and the structural information image are respectively substituted into a single residual block to obtain a basic feature and a structural basic feature; The basic feature and the structural basic feature are substituted into a multi-scale sampling model to obtain a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature; The low-scale morphological feature, the medium-scale morphological feature and the high-scale morphological feature are substituted into a double-path feature optimization model to obtain a low-scale final fusion feature, a medium-scale final fusion feature and a high-scale final fusion feature; According to the low-scale final fusion feature, the medium-scale final fusion feature and the high-scale final fusion feature, a target high-quality sperm is screened in combination with a morphological quality standard.

6. The high precision sperm morphology screening method according to claim 5, wherein, The second initial single-layer active sperm image and the structural information image are respectively substituted into a single residual block to obtain a basic feature and a structural basic feature, including: The second initial single-layer active sperm image and the structural information image are respectively substituted into a single residual block to obtain a first feature and a second feature; The first feature and the second feature are compressed in feature through average pooling and maximum pooling, and are merged to obtain a fusion feature; The fusion feature, the first feature and the second feature are spliced to obtain a joint feature; The joint feature is subjected to a 5x5 convolution layer plus Sigmoid activation to generate a first weight and a second weight; the first weight is multiplied by the first feature to obtain a basic feature; and the second weight is multiplied by the second feature to obtain a structural basic feature.

7. The method of claim 5, wherein the high-precision sperm morphology screening method is characterized by, The basic feature and the structural basic feature are substituted into a multi-scale sampling model to obtain a low-scale morphological feature, a medium-scale morphological feature and a high-scale morphological feature, including: The basic feature and the structural basic feature are fused to obtain a fusion basic feature; The fusion basic feature is split into a first sub-fusion basic feature and a second sub-fusion basic feature through a 1x1 convolution; The first sub-fusion basic feature is subjected to convolution operation through a first convolution kernel to obtain a first sub-convolution feature; and the second sub-fusion basic feature is subjected to convolution operation through a second convolution kernel to obtain a second sub-convolution feature; The second sub-convolution feature is subjected to 3x3 convolution plus 1x1 convolution plus Sigmoid activation to obtain a second sub-convolution attention; The second sub-convolution attention is multiplied by the second sub-convolution feature to obtain a second sub-attention feature; The second sub-attention feature and the first sub-convolution feature are spliced, and then a 3*3 convolution and a residual connection are used to obtain an enhanced multi-stage feature; the multi-stage feature includes a low-scale morphological feature, a medium-scale morphological feature, and a high-scale morphological feature.

8. The high precision sperm morphology screening method according to claim 5, wherein, The low-scale morphological feature, the medium-scale morphological feature, and the high-scale morphological feature are input into a double-path feature optimization model to obtain a low-scale final fusion feature, a medium-scale final fusion feature, and a high-scale final fusion feature, including: The first scale feature and the second scale feature are first subjected to a 1*1 convolution, and then the second scale feature is up-sampled, and the first scale feature and the second scale feature are fused and then subjected to bilinear interpolation to generate a first semantic feature; the first scale feature is any one of the medium-scale morphological feature and the high-scale morphological feature; the second scale feature is any one of the medium-scale morphological feature and the low-scale morphological feature; the scale of the second scale feature is smaller than that of the first scale feature; The first semantic feature is subjected to a 1*1 convolution and then input into a Sigmoid activation function to obtain a foreground attention map and a background attention map; The foreground attention map is multiplied by the second scale feature to obtain a foreground path, and the background attention map is multiplied by the second scale feature to obtain a background path; when the first scale feature is the high-scale morphological feature and the second scale feature is the medium-scale morphological feature, a first foreground path and a first background path are obtained; when the first scale feature is the high-scale morphological feature and the second scale feature is the low-scale morphological feature, a second foreground path and a second background path are obtained; when the first scale feature is the medium-scale morphological feature and the second scale feature is the low-scale morphological feature, a third foreground path and a third background path are obtained; The first foreground path, the second foreground path, and the third foreground path are fused to obtain a fused foreground path; The first background path, the second background path, and the third background path are fused to obtain a fused background path; The fused foreground path and the fused background path are respectively optimized by a residual block, spliced, and then subjected to a 1*1 convolution to generate a final fusion feature; The low-scale morphological feature and the final fusion feature are averaged to obtain a low-scale final fusion feature; the medium-scale morphological feature and the final fusion feature are averaged to obtain a medium-scale final fusion feature; and the high-scale morphological feature and the final fusion feature are averaged to obtain a high-scale final fusion feature.

9. The method of claim 1, wherein the method is a high precision method of sperm morphology screening. The first magnification operation is implemented by using a microscope; the microscope uses a 60x objective lens and a 10x ocular lens.

10. A high precision sperm morphology screening system, characterized by, The system comprises: A filtering module configured to obtain raw semen, and filter the raw semen to obtain initial semen; An initial single-layer active sperm screening module configured to spread the initial semen into a preset single-layer sperm liquid strip to obtain initial single-layer active sperm; A first magnification module configured to perform a first magnification operation on the initial single-layer active sperm to obtain a first initial single-layer active sperm image; A magnified initial single-layer active sperm image generation module configured to magnify the first initial single-layer active sperm image to a target resolution to obtain a magnified initial single-layer active sperm image; and A single-layer active sperm image screening module configured to screen the magnified initial single-layer active sperm image to obtain a single-layer active sperm image. a second amplification module, configured to perform a second amplification operation on the first initial single-layer active sperm image according to the amplified initial single-layer active sperm image to obtain a second initial single-layer active sperm image; a structural information image generation module, configured to obtain a structural information image by using a Sobel operator for the second initial single-layer active sperm image; a high-precision sperm morphology selection module, configured to obtain a target high-quality sperm by substituting the second initial single-layer active sperm image and the structural information image into a high-precision sperm morphology selection model.

Images