An antral follicle recognition method, system and electronic device based on super-resolution

Abstract

This invention provides a method, system, electronic device, and storage medium for antral follicle identification based on super-resolution. The method includes: acquiring ovarian ultrasound images; using a target detection model to identify and mask out larger first-type follicle regions; magnifying the obtained preprocessed image and enhancing it using a super-resolution generative adversarial network to improve image details; inputting the enhanced image into a semantic segmentation model to obtain a pixel-level segmentation mask for smaller second-type follicles; and finally mapping the segmentation mask back to the original image coordinate space to obtain the final segmentation result. This invention can effectively eliminate interference from large follicles, enhance the features of small antral follicles, and significantly improve the accuracy and reliability of segmentation and identification, providing accurate quantitative evidence for ovarian function assessment.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of medical image processing technology, and in particular to a method and system for identifying antral follicles based on super-resolution. Background Technology

[0002] In the fields of assisted reproduction and gynecological endocrinology, accurate identification and counting of antral follicles in the ovary using ultrasound images is a crucial step in assessing ovarian reserve. Antral follicles are follicles in the early stages of development, with small diameters, typically between 2-9 mm. Due to their small size, limited pixel information in ultrasound images, indistinct edges, and frequent mixing with other ovarian tissues (such as blood vessels and stroma) and larger mature follicles, automated segmentation and identification of antral follicles present significant challenges. Current automated follicle segmentation methods often directly perform end-to-end semantic segmentation or target detection on the entire ultrasound image. However, while these methods often perform well on larger, well-defined mature follicles, their segmentation accuracy drops significantly for small, indistinct antral follicles, leading to frequent missed detections (failure to identify existing antral follicles) and false detections (identifying non-follicular tissue as antral follicles). This severely impacts the accuracy and reliability of clinical assessments based on antral follicle counting.

[0003] In recent years, super-resolution technology has been introduced into the field of medical image processing, aiming to improve the resolution and clarity of images through algorithms. However, directly applying super-resolution technology to raw ultrasound images containing follicles of various sizes, while improving overall detail, can lead to the strong features of large, mature follicles overshadowing or interfering with the enhancement and focus on the subtle features of antral follicles. Furthermore, computational resources may be heavily consumed in non-critical areas. Therefore, the key challenge in improving the accuracy of antral follicle identification lies in how to selectively enhance the features of the antral follicle region while suppressing or eliminating interference from mature follicles. Summary of the Invention

[0004] This invention provides a super-resolution-based method, system, electronic device, and storage medium for identifying antral follicles, addressing the technical problems of inaccurate segmentation and identification of small, weakly characterized antral follicles in ultrasound images, resulting in high rates of missed detection and false detection.

[0005] In a first aspect, embodiments of the present invention provide a method for identifying antral follicles based on super-resolution, comprising: S1. Obtain the original ultrasound image of the ovarian region, and detect the original ultrasound image based on the preset target detection model to identify the first type of follicle region with a size greater than or equal to the first threshold.

[0006] S2. Based on the identified first type of follicle region, generate a corresponding mask; apply the mask to process the original ultrasound image to eliminate the display of the first type of follicle region, and obtain a first preprocessed image.

[0007] S3. Perform image magnification processing on the first preprocessed image to obtain a magnified image; input the magnified image into a super-resolution generative adversarial network to generate a high-resolution image.

[0008] S4. Input the high-resolution image into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask for the second type of follicle in the high-resolution image; wherein, the second type of follicle is a follicle with a size smaller than the first threshold.

[0009] S5. Map the pixel-level segmentation mask back to the coordinate space of the original ultrasound image to generate a final antral follicle segmentation map that corresponds one-to-one with the original ultrasound image in spatial location.

[0010] Preferably, in step S1, the preset target detection model is a lightweight target detection network with single-channel input.

[0011] The preset target detection model is obtained through training. The training set used for training is constructed by performing data augmentation processing on the original ultrasound images labeled with the first type of follicle bounding boxes. The data augmentation processing includes at least one of grayscale adjustment, random rotation, and Gaussian noise injection.

[0012] Preferably, in step S2, the mask is applied to process the original ultrasound image to eliminate the display of the first type of follicle region, including: The pixel values ​​corresponding to the first type of follicle region in the original ultrasound image are set to a preset background grayscale value.

[0013] Preferably, in step S3, the generator of the super-resolution generative adversarial network integrates a Transformer module; the training loss function of the super-resolution generative adversarial network includes a perceptual loss function and an edge-preserving loss function.

[0014] Preferably, in step S4, the semantic segmentation model is a U-Net network structure, and the training process of the semantic segmentation model includes: The training image is input into the semantic segmentation model to obtain the initial prediction result.

[0015] The initial prediction results are then subjected to morphological filtering and connected component analysis to remove artifact regions, resulting in optimized prediction results.

[0016] The loss is calculated based on the difference between the optimized prediction result and the true label, and the model parameters are updated based on the loss calculated based on the difference.

[0017] Preferably, the first threshold is 5 mm, and the second type of follicle is an antral follicle.

[0018] Preferably, in step S5, mapping the pixel-level segmentation mask back to the coordinate space of the original ultrasound image includes: The pixel-level segmentation mask is downsampled to obtain an initial restored segmentation image; The segmentation region corresponding to the second type of follicle in the initial restored segmentation map is subjected to edge refinement processing to generate the final antral follicle segmentation map; wherein, the small-sized target.

[0019] Secondly, the present invention provides a super-resolution-based antral follicle identification system, comprising: The first detection module is used to acquire the original ultrasound image of the ovarian region, and to detect the original ultrasound image based on a preset target detection model to identify the first type of follicle region whose size is greater than or equal to a first threshold.

[0020] The preprocessing module is used to generate a corresponding mask based on the identified first type of follicle region; and to process the original ultrasound image using the mask to eliminate the display of the first type of follicle region, thereby obtaining a first preprocessed image.

[0021] The super-resolution reconstruction module is used to perform image magnification processing on the first preprocessed image to obtain a magnified image; the magnified image is then input into a super-resolution generative adversarial network to generate a high-resolution image.

[0022] The second segmentation module is used to input the high-resolution image into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask of the second type of follicle in the high-resolution image; wherein the second type of follicle is a follicle with a size smaller than the first threshold.

[0023] The coordinate mapping and output module is used to map the pixel-level segmentation mask back to the coordinate space of the original ultrasound image, generating a final antral follicle segmentation map that corresponds one-to-one with the original ultrasound image in spatial location.

[0024] Thirdly, embodiments of the present invention provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the super-resolution-based antral follicle identification method as described in the first aspect of the present invention.

[0025] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the super-resolution-based antral follicle identification method as described in the first aspect of the present invention.

[0026] This invention provides a super-resolution-based antral follicle identification method, system, electronic device, and storage medium. Through a multi-stage progressive processing flow, it specifically addresses the difficulty of identifying small-sized antral follicles in ultrasound images. First, a target detection model is used to locate and mask out larger, more prominent mature follicles in the image, eliminating their interference with subsequent steps. Next, the preprocessed image containing only small targets and background is directionally magnified, and a super-resolution generative adversarial network is used to enhance its resolution and detail features, highlighting the subtle edges and morphological information of the antral follicles. Then, a semantic segmentation model is used to perform pixel-level fine segmentation of smaller second-type follicles on the enhanced image. Finally, the segmentation results are precisely mapped back to the original image coordinate space, outputting a final segmentation result that perfectly matches the spatial position of the original ultrasound image. It significantly improves the overall accuracy of antral follicle segmentation and identification, effectively solving the problems of high false negative and false positive rates caused by small target size, few pixels, blurred edges, and interference from large follicles. The targeted application of super-resolution technology improves the image quality of the antral follicle region, enabling subsequent segmentation to capture clearer edge details, thereby obtaining segmentation boundaries that better fit the real anatomical structure. At the same time, the strict spatial correspondence between the final result and the original image ensures the clinical usability of the segmentation results, providing a reliable and accurate quantitative basis for medical analysis such as ovarian reserve function assessment. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of a super-resolution-based antral follicle identification method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of follicle image changes according to an embodiment of the present invention; Figure 3 This is a schematic diagram of super-resolution generative adversarial network processing according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the U-Net network structure training according to an embodiment of the present invention; Figure 5This is a block diagram of a super-resolution-based antral follicle recognition system according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the physical structure according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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 scope of protection of the present invention.

[0030] In the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] This invention provides a super-resolution-based method for identifying antral follicles, such as... Figure 1 , Figure 2 As shown, it includes: S1. Obtain the original ultrasound image of the ovarian region, and detect the original ultrasound image based on the preset target detection model to identify the first type of follicle region with a size greater than or equal to the first threshold.

[0033] The original ultrasound image refers to the grayscale image of the ovarian region acquired by ultrasound equipment in clinical practice. It is the basic data for antral follicle identification and must meet the requirements of clinical compliance and a resolution of ≥512×512 pixels to be compatible with subsequent model processing. The preset target detection model is an artificial intelligence network that has been pre-trained and optimized and is specifically used to identify follicles of specific sizes in ultrasound images. In this embodiment, the YOLOv8-nano lightweight framework adapted to ultrasound grayscale images is specifically used. The first threshold is the critical value for distinguishing follicles of different sizes, which is set to 5mm in combination with the clinical definition of antral follicles and detection requirements. The first type of follicle region is the region where normal follicles with a diameter of ≥5mm are located. These follicles are larger in size, have obvious features, and are easy to identify in ultrasound images. Existing technologies do not specifically separate large-sized normal follicles from antral follicles, causing the characteristics of normal follicles to mask antral follicles and interfere with subsequent identification. This step uses a target detection model to accurately identify the first type of follicle region. The core logic is to first separate easily identifiable large-sized follicles, solving the problem of interference from normal follicles in antral follicle identification. This provides a precise target region reference for subsequent interference elimination and focused antral follicle processing. At the same time, the model is trained with clinically compliant data and optimized in multiple dimensions to ensure the accuracy of the first type of follicle region identification. The validation set mAP (Mean Average Precision) target is ≥90%, and the false negative rate is ≤5%.

[0034] S2. Based on the identified first type of follicle region, generate a corresponding mask; apply the mask to process the original ultrasound image to eliminate the display of the first type of follicle region, and obtain a first preprocessed image.

[0035] It should be noted that a mask is an image processing tool used to obscure specific areas of an image. In this embodiment, the target area is generated by setting the pixel value to 255, thus achieving the masking of a specific area. The first preprocessed image is an intermediate image obtained after masking. Its core feature is the elimination of the display of the first type of follicle area, retaining only the antral follicle and background tissues such as ovarian vessels and stroma. Existing technologies do not eliminate the interference of normal follicles on antral follicles, causing antral follicles to be easily covered or confused by normal follicles in the original ultrasound image, making it difficult for subsequent models to accurately capture them. This step generates a mask based on the first type of follicle area identified by S1 and processes the original ultrasound image. Logically, it achieves the initial highlighting of the antral follicle area by shielding the interference source, solving the technical pain point of mutual interference between normal follicles and antral follicles in the image, obtaining image data that focuses only on antral follicles and background tissues, significantly reducing environmental interference in subsequent antral follicle identification, and improving the model's attention to antral follicle features.

[0036] S3. Perform image magnification processing on the first preprocessed image to obtain a magnified image; input the magnified image into a super-resolution generative adversarial network to generate a high-resolution image.

[0037] Image magnification is a preprocessing operation that enlarges the image size by a fixed ratio. In this embodiment, the magnification factor is set to 4 times to increase the pixel ratio of antral follicles in the image. The magnified image, obtained after 4x magnification, has an increased size but no improvement in detail clarity, requiring further optimization. Super-Resolution Generative Adversarial Network (SRGAN) is an artificial intelligence network that improves image resolution and detail performance through adversarial training between the generator and the discriminator. In this embodiment, the generator is improved by adding a transformer module to enhance feature extraction capabilities. The high-resolution image is an image that has been processed by super-resolution technology, resulting in enhanced edge details and feature information, which can clearly present the morphological features of antral follicles. The core drawback of existing technologies is the low resolution of ultrasound images, the small size of antral follicles, the low pixel ratio, and the blurred edges, making it difficult for their features to be recognized by segmentation models. This step first enlarges the image and then processes it through an improved super-resolution network. Logically, it first increases the pixel ratio and then enhances the feature expression, which solves the key technical problems of unclear antral follicle features and missing details. This significantly enhances the core features of antral follicles, such as their edges and morphology, providing high-quality image data support for subsequent pixel-level segmentation.

[0038] S4. Input the high-resolution image into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask for the second type of follicle in the high-resolution image; wherein, the second type of follicle is a follicle with a size smaller than the first threshold.

[0039] The semantic segmentation model is an artificial intelligence network capable of classifying each pixel in an image. In this embodiment, an improved U-Net network is used, combined with morphological filtering and connected component analysis (Laplacian algorithm) to optimize the training process. The pixel-level segmentation mask is a binary image output by the semantic segmentation model that accurately marks the position of each pixel of the second type of follicle, providing a direct representation of the spatial location of the antral follicle. The second type of follicle, namely the antral follicle with a diameter smaller than the first threshold (5mm), is the core identification target in this embodiment. Due to its small size and susceptibility to background interference, it needs to be accurately segmented for confirmation. Existing semantic segmentation models are not optimized for the features of antral follicles and do not handle pseudo-detection regions, resulting in low segmentation accuracy and high false positive and false negative rates. This step inputs a high-resolution image into a specially optimized semantic segmentation model. Logically, it performs accurate pixel segmentation based on the enhanced antral follicle features, solving the technical problems of inaccurate antral follicle segmentation and interference from pseudo-detection regions. This achieves accurate pixel-level segmentation of antral follicles, effectively filters pseudo-detection regions, significantly reduces the false negative and false positive rates, and obtains accurate segmentation results of antral follicles in magnified images.

[0040] S5. Map the pixel-level segmentation mask back to the coordinate space of the original ultrasound image to generate a final antral follicle segmentation map that corresponds one-to-one with the original ultrasound image in spatial location.

[0041] The process of mapping back to the coordinate space of the original ultrasound image refers to reducing the pixel-level segmentation mask by a factor of 4 (the magnification ratio in S3) to ensure that the size and spatial position of the segmentation result are consistent with the original ultrasound image. The final antral follicle segmentation map is a result that perfectly corresponds to the original ultrasound image in terms of spatial position and scale, and can be directly used for relevant clinical medical assessments. In existing technologies, some segmentation schemes have poor compatibility with the original image, resulting in inconsistent spatial positions or scales, failing to provide effective reference for clinical diagnosis. This step maps the segmentation result in the magnified image back to the coordinate space of the original image. Logically, this achieves a precise match between the processed result and the original data, solving the technical problem of poor spatial compatibility between the segmentation result and the original ultrasound image. This ensures that the antral follicle segmentation result accurately corresponds to the actual position and morphology in the original ultrasound image, providing reliable and usable quantitative information on antral follicles for clinical ovarian reserve function assessment and other medical applications, thus guaranteeing the accuracy of diagnostic assessments.

[0042] Based on the above embodiments, as a preferred implementation, in step S1, the preset target detection model is a lightweight target detection network with single-channel input; The preset target detection model is obtained through training. The training set used for training is constructed by performing data augmentation processing on the original ultrasound images labeled with the first type of follicle bounding boxes. The data augmentation processing includes at least one of grayscale adjustment, random rotation, and Gaussian noise injection.

[0043] The lightweight object detection network refers to an AI detection network with a simple model structure, low computational cost, and fast inference speed, adapted to the operating environment of clinical ultrasound equipment. In this implementation, the YOLOv8-nano framework is specifically adopted, which can improve processing efficiency while ensuring detection accuracy. Single-channel input addresses the characteristic of ultrasound images being grayscale images by changing the network's default three-channel input to a single channel, making the network more compatible with the data format of ultrasound images and reducing redundant computation. The training set is a sample set used for model training, constructed from clinically compliant ovarian ultrasound images, requiring a quantity ≥300 cases and a single image pixel ≥512×512 to ensure the representativeness and sufficiency of the samples. The bounding boxes of the first type of follicles indicate... Axis-aligned bounding boxes are used to mark the locations of first-type follicles with a diameter ≥5mm in ultrasound images, and the annotation results must be reviewed and confirmed by two people to ensure the accuracy and reliability of the annotation. Data augmentation is a series of image processing techniques to expand the diversity of the training set and improve the generalization ability of the model. Gray-scale adjustment refers to adjusting the gray-scale value of the image within a range of ±20% to simulate the gray-scale changes of ultrasound images under different imaging conditions. Random rotation refers to randomly rotating the image within a range of -15° to 15° to adapt to the different angles that follicles may appear in ultrasound images. Gaussian noise injection refers to adding Gaussian noise with a variance of 0.01-0.03 to the image to simulate the noise interference that may be generated during ultrasound imaging. This embodiment first determines a lightweight network adapted to ultrasound images and adjusts the input channels. Then, a high-quality training set is constructed through standardized sample annotation and multi-dimensional data augmentation. Combined with reasonable training parameters (batch size 8, initial learning rate 1e-4, AdamW optimizer, combination of CIoU loss and class loss, 300 iterations, validation set percentage 20%, early termination if mAP does not increase for 10 consecutive iterations) and full-process monitoring (tracking training set loss, validation set mAP ≥ 90%, false negative rate ≤ 5%), the model training is completed. After training, the generalization ability is evaluated using a 10% test set. If insufficient, data is added or the learning rate is fine-tuned to obtain the optimal model. This solves the technical problems of poor adaptability, low training set quality, and insufficient generalization ability of existing models, which lead to inaccurate identification of the first type of follicle. It achieves accurate and efficient identification of the first type of follicle, providing a reliable foundation for eliminating its interference with antral follicles. At the same time, the lightweight characteristics of the model are suitable for clinical application scenarios, ensuring that the detection efficiency can meet the actual needs of clinical diagnosis and treatment.

[0044] Based on the above embodiments, as a preferred implementation, in step S2, the original ultrasound image is processed using the mask to eliminate the display of the first type of follicle region, including: The pixel values ​​corresponding to the first type of follicle region in the original ultrasound image are set to a preset background grayscale value.

[0045] Specifically, the preset background grayscale value is a unified grayscale value pre-set based on the imaging characteristics of ultrasound grayscale images and subsequent image processing requirements. In the clinical antral follicle recognition scenario, it is specifically set to 255. This value corresponds to the white background in the ultrasound image, which can effectively occlude the target area. The pixels corresponding to the first type of follicle area refer to all pixels in the target frame of the first type of follicle with a diameter ≥5mm identified by the target detection model in the original ultrasound image. These pixels carry the morphological, grayscale and other feature information of the first type of follicle, which will interfere with the recognition of antral follicles. In existing technologies, some follicle identification schemes do not use precise pixel value replacement to eliminate interference from large-sized follicles, or only use simple cropping and blurring. This results in residual features of the first type of follicles still affecting the subsequent identification of antral follicles and may damage the overall integrity of the image. This embodiment, based on the precise positioning of the first type of follicle region in step S1, constructs a targeted mask by uniformly setting all pixels in this region to the background grayscale value of 255 and applying it to the original ultrasound image. This achieves complete shielding of the first type of follicle region, solving the technical defects of incomplete elimination of first type of follicle interference and easy residual feature interference in existing technologies. It generates a first preprocessed image that only retains the background tissues such as antral follicles, ovarian blood vessels, and stroma. This not only ensures the integrity of the overall image structure, but also removes the main interference sources for antral follicle identification to the greatest extent, allowing subsequent super-resolution processing and semantic segmentation to focus more on the weak features of antral follicles, laying the foundation for improving the accuracy of antral follicle identification.

[0046] Based on the above embodiments, as a preferred implementation, in step S3, such as Figure 3 As shown, the generator of the super-resolution generative adversarial network integrates a Transformer module; the training loss function of the super-resolution generative adversarial network includes a perceptual loss function and an edge-preserving loss function.

[0047] Among them, the Transformer module is a neural network component that can capture the global feature correlation of an image. When integrated into the generator, it can enhance the ability to extract weak features of antral follicles, improve the accuracy of feature extraction, and adapt to the characteristics of small antral follicle size and scattered features. The generator of the super-resolution generative adversarial network is the core execution unit of super-resolution technology. It is built based on SRGAN (Super-Resolution Generative Adversarial Network) and is responsible for transforming low-quality magnified images into high-resolution images, which is the key to achieving image detail enhancement.

[0048] Furthermore, the perceptual loss function is a loss calculation method that measures the difference in features between the network output image and the real high-resolution image. It encourages the output image to maintain consistency with the real image in terms of content structure, without pursuing pixel-level perfect matching, thus adapting to the need for antral follicle feature enhancement. In this embodiment, the perceptual loss function adopts the Feature Reconstruction Loss method to measure the super-resolution image output by the super-resolution generative adversarial network. Compared with real super-resolution images y (That is, ovarian ultrasound images acquired by high-resolution clinical ultrasound equipment are used as real labels for network training to ensure clinical suitability) in the pre-trained loss network Specific feature layers j The above represents the method for calculating the loss due to the difference:

[0049] in, It is a loss network Feature layer j The activation (feature map) of the network is used to extract abstract features from images (adapting to the feature extraction requirements of antral follicle ultrasound images), and the feature layer is used for this purpose. j The shape is C j × H j × W j ( C j , H j , W j These represent the number of channels, height, and width of the layer, respectively, ensuring consistency in feature structure between the super-resolution image and the real image.

[0050] Furthermore, the edge-preserving loss function is a loss function for image edge regions, where E is the pre-extracted antral follicle edge features (adapting to the boundary recognition requirements between antral follicles and background tissue). By weighting the absolute error of the edge region, the edge details of the antral follicles are enhanced. The edge loss function in this embodiment is:

[0051] Among them, boundary information extracted based on clinically labeled antral follicle segmentation results is obtained through edge features E, N E The total number of pixels for the edge feature E is used to enhance the boundary distinction between the antral follicle and the background tissue. However, this embodiment is more concerned with the edge information, so E is used for weighting, which increases the proportion of the edge part. This allows the loss to better optimize the edge information.

[0052]

[0053]

[0054] The total loss (loss_total) is a weighted combination of pixel loss and edge loss, where the weighting coefficient α is used to balance the optimization weights of the two. In this invention, α is set to 0.3 (emphasizing edge detail enhancement) to balance the overall fidelity of the super-resolution image with the clarity of the antral follicle edge.

[0055] Based on the above embodiments, as a preferred implementation method, such as Figure 4 As shown in the diagram, in step S4, the semantic segmentation model is a U-Net network structure, and the training process of the semantic segmentation model includes: The training image is input into the semantic segmentation model to obtain the initial prediction result.

[0056] The initial prediction results are then subjected to morphological filtering and connected component analysis to remove artifact regions, resulting in optimized prediction results.

[0057] The loss is calculated based on the difference between the optimized prediction result and the true label, and the model parameters are updated based on the loss calculated based on the difference.

[0058] This implementation employs an improved architecture, integrating CNN local feature extraction and Transformer global context modeling capabilities to adapt to the segmentation requirements of small-sized, low-percentage antral follicles. Training images refer to high-resolution ovarian ultrasound images output by S3, with antral follicle pixel-level locations annotated by clinicians and double-checked to construct a training set containing ≥300 samples. Each image has a pixel size ≥2048×2048 (after 4x magnification) to ensure sample representativeness. The initial prediction result is the model's preliminary pixel-level classification of antral follicles in the training images, without artifact removal processing and potentially containing false positives. Morphological filtering is an image processing technique based on mathematical morphology, using flat structuring elements to filter the initial prediction result, removing artifacts caused by poor probe contact in ultrasound imaging. Point-like and short-line noise artifacts caused by sound beam attenuation; connected component analysis extracts edge features using the Laplacian algorithm, filters connected regions with areas within a preset threshold (adapting to the actual size of antral follicles), and removes artifact connected regions with excessively small areas; artifact regions refer to interference areas in ultrasound images unrelated to antral follicles, including reverberation artifacts and partial volume artifacts, which are easily misidentified as antral follicles, affecting segmentation accuracy; the optimized prediction result is an antral follicle prediction mask after morphological filtering and connected component analysis to remove artifacts, which is closer to the actual distribution of antral follicles; the true label is the pixel-level annotation result of antral follicles in the training image, serving as the benchmark for model training; model parameters refer to the learnable parameters of the U-Net network, such as convolutional kernel weights and biases, which are updated through backpropagation of the loss. This embodiment first inputs labeled high-resolution training images into an improved U-Net network (integrating CNN local feature extraction and Transformer global feature modeling to enhance small target recognition capabilities), outputting initial prediction results. Then, morphological filtering smooths noise artifacts, and connected component analysis using Laplacian edge extraction precisely removes artifact regions, yielding optimized prediction results. Finally, the Dice loss function (adapting to class imbalance) is used to calculate the difference between the optimized prediction results and the true labels. Backpropagation updates the model parameters, and iterative training continues until the loss converges (Dice coefficient ≥ 0.85 on the validation set). This solves the technical problems of low segmentation accuracy caused by traditional semantic segmentation models' failure to handle ultrasound artifacts and insufficient adaptability. It significantly reduces the interference of artifacts on antral follicle segmentation, improves the accuracy of antral follicle edge localization and the integrity of the segmentation mask, effectively alleviates model bias caused by class imbalance, and enables the semantic segmentation model to accurately identify antral follicles in high-resolution images. This provides a reliable segmentation basis for subsequent mapping back to the original image coordinate space, further ensuring the accuracy of clinical antral follicle counting and morphological assessment.

[0059] Based on the above embodiments, as a preferred implementation, step S5, mapping the pixel-level segmentation mask back to the coordinate space of the original ultrasound image, includes: The pixel-level segmentation mask is downsampled to obtain an initial restored segmentation map.

[0060] The segmentation region corresponding to the second type of follicle in the initial restored segmentation map is subjected to edge refinement processing to generate the final antral follicle segmentation map; wherein, the small-sized target.

[0061] The downsampling process involves reducing the pixel-level segmentation mask by a factor of four, consistent with the magnification of the S3 image. It employs a nearest-neighbor interpolation algorithm to avoid introducing false pixel information, ensuring that the spatial location of the antral follicle segmentation region accurately corresponds to the original ultrasound image, thus meeting the accuracy requirements for label restoration in medical image segmentation. The initial restored segmentation image, after downsampling, is a segmentation image with a size completely identical to the original ultrasound image (≥512×512 pixels), achieving preliminary spatial matching, but may contain jagged redundant pixels at the edges. Edge refinement is a composite process combining mathematical morphology. The process involves first shrinking redundant pixels at the edges of the segmented region using erosion operations on circular structuring elements, and then employing a skeleton refinement algorithm to extract precise edges for each pixel, adapting to the small size and easily distorted edges of antral follicles. The final antral follicle segmentation image, after downsampling and edge refinement optimization, is a precise segmentation result that perfectly matches the original ultrasound image in terms of spatial location, size, and target morphology, and can be directly used for clinical antral follicle counting and morphological assessment. The second type of follicle, namely antral follicles with a diameter less than 5mm, is the core optimization target of this embodiment. Its small size requires extremely high spatial accuracy and edge clarity in the segmentation result. In existing technologies, traditional segmentation mask mapping simply scales the size without optimizing for the small size of antral follicles and the blurred edges of ultrasound images. This results in reconstructed segmentation images with rough edges, obvious jagged edges, and slight spatial positional shifts, failing to accurately reflect the actual morphology and location of the antral follicles in the original image, thus affecting the accuracy of clinical diagnosis. This embodiment first uses a nearest-neighbor interpolation algorithm to perform a 4x downsampling process on the pixel-level segmentation mask based on the 4x image magnification ratio in S3, restoring its size from the magnified ≥2048×2048 pixels to match the original ultrasound image, thus obtaining an initial restored segmentation map. This solves the problems of size mismatch and spatial misalignment between the segmentation mask and the original image. Next, to address the edge redundancy and jaggedness defects in the antral follicle region of the initial restored segmentation map, the edges are smoothed using erosion operations on circular structuring elements to remove false redundant pixels. Subsequently, a skeleton refinement algorithm is used to extract precise single-pixel edges, achieving fine-grained optimization of the antral follicle edges, and finally generating the final antral follicle segmentation map. This overcomes the technical shortcomings of traditional mapping methods, such as rough edges and insufficient positional accuracy, ensuring that the segmentation result is spatially aligned with the original ultrasound image at the sub-pixel level. The antral follicle edges are clear and smooth, and the morphology closely matches reality, effectively improving the accuracy of clinical antral follicle counting and the reliability of morphological assessment, providing high-quality quantitative evidence for ovarian reserve function diagnosis.

[0062] Secondly, the present invention provides a super-resolution-based antral follicle identification system, based on the super-resolution-based antral follicle identification methods described in the above embodiments, such as... Figure 5 As shown, the system includes: The first detection module 510 is used to acquire the original ultrasound image of the ovarian region, detect the original ultrasound image based on a preset target detection model, and identify the first type of follicle region whose size is greater than or equal to a first threshold. The preprocessing module 520 is used to generate a corresponding mask based on the identified first type of follicle region; and to process the original ultrasound image using the mask to eliminate the display of the first type of follicle region to obtain a first preprocessed image. The super-resolution reconstruction module 530 is used to perform image magnification processing on the first preprocessed image to obtain a magnified image; and input the magnified image into a super-resolution generative adversarial network to generate a high-resolution image. The second segmentation module 540 is used to input the high-resolution image into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask of the second type of follicle in the high-resolution image; wherein, the second type of follicle is a follicle with a size smaller than the first threshold. The coordinate mapping and output module 550 is used to map the pixel-level segmentation mask back to the coordinate space of the original ultrasound image, generating a final antral follicle segmentation map that corresponds one-to-one with the original ultrasound image in spatial location.

[0063] Based on the same concept, this invention also provides a schematic diagram of a physical structure, such as... Figure 6 As shown, the server may include a processor 610, a communications interface 620, a memory 730, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions stored in the memory 630 to execute the steps of the super-resolution-based antral follicle identification method as described in the above embodiments.

[0064] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] Based on the same concept, embodiments of the present invention also provide a non-transitory computer-readable storage medium storing a computer program containing at least one piece of code that can be executed by a master control device to control the master control device to implement the steps of the super-resolution-based antral follicle identification method as described in the above embodiments.

[0066] Based on the same technical concept, this application also provides a computer program, which, when executed by a main control device, is used to implement the above-described method embodiments.

[0067] The program may be stored, in whole or in part, on a storage medium packaged with the processor, or in part or in whole on a memory not packaged with the processor.

[0068] Based on the same technical concept, embodiments of this application also provide a processor for implementing the above-described method embodiments. The processor may be a chip.

[0069] The various embodiments of the present invention can be combined arbitrarily to achieve different technical effects.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying antral follicles based on super-resolution, characterized in that, include: S1. Obtain the original ultrasound image of the ovarian region, and detect the original ultrasound image based on the preset target detection model to identify the first type of follicle region with a size greater than or equal to the first threshold. S2. Based on the identified first type of follicle region, generate a corresponding mask; The original ultrasound image is processed using the mask to eliminate the display of the first type of follicle region, resulting in a first preprocessed image; S3. Perform image magnification processing on the first preprocessed image to obtain a magnified image; The magnified image is input into a super-resolution generative adversarial network to generate a high-resolution image; S4. Input the high-resolution image into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask for the second type of follicles in the high-resolution image; wherein, the second type of follicles are follicles with a size smaller than the first threshold. S5. Map the pixel-level segmentation mask back to the coordinate space of the original ultrasound image to generate a final antral follicle segmentation map that corresponds one-to-one with the original ultrasound image in spatial location.

2. The antral follicle identification method based on super-resolution according to claim 1, characterized in that, In S1, the preset target detection model is a lightweight target detection network with single-channel input; The preset target detection model is obtained through training. The training set used for training is constructed by performing data augmentation processing on the original ultrasound images labeled with the first type of follicle bounding boxes. The data augmentation processing includes at least one of grayscale adjustment, random rotation, and Gaussian noise injection.

3. The antral follicle identification method based on super-resolution according to claim 1, characterized in that, In step S2, the mask is applied to process the original ultrasound image to eliminate the display of the first type of follicle region, including: The pixel values ​​corresponding to the first type of follicle region in the original ultrasound image are set to a preset background grayscale value.

4. The antral follicle identification method based on super-resolution according to claim 1, characterized in that, In S3, the generator of the super-resolution generative adversarial network integrates a Transformer module; the training loss function of the super-resolution generative adversarial network includes a perceptual loss function and an edge-preserving loss function.

5. The method for identifying antral follicles based on super-resolution according to claim 1, characterized in that, In step S4, the semantic segmentation model is a U-Net network structure, and the training process of the semantic segmentation model includes: The training image is input into the semantic segmentation model to obtain the initial prediction result; The initial prediction results are sequentially subjected to morphological filtering and connected component analysis to remove artifact regions and obtain optimized prediction results. The loss is calculated based on the difference between the optimized prediction result and the true label, and the model parameters are updated based on the loss calculated based on the difference.

6. The method for identifying antral follicles based on super-resolution according to claim 1, characterized in that, The first threshold is 5 mm, and the second type of follicle is an antral follicle.

7. The antral follicle identification method based on super-resolution according to claim 1, characterized in that, In step S5, mapping the pixel-level segmentation mask back to the coordinate space of the original ultrasound image includes: The pixel-level segmentation mask is downsampled to obtain an initial restored segmentation image; The segmentation region corresponding to the second type of follicle in the initial restored segmentation map is subjected to edge refinement processing to generate the final antral follicle segmentation map; wherein, the small-sized target.

8. A super-resolution-based antral follicle identification system, characterized in that, include: The first detection module is used to acquire the original ultrasound image of the ovarian region, detect the original ultrasound image based on a preset target detection model, and identify the first type of follicle region whose size is greater than or equal to a first threshold. The preprocessing module is used to generate a corresponding mask based on the identified first type of follicle region; The original ultrasound image is processed using the mask to eliminate the display of the first type of follicle region, resulting in a first preprocessed image; The super-resolution reconstruction module is used to perform image magnification processing on the first preprocessed image to obtain a magnified image; The magnified image is input into a super-resolution generative adversarial network to generate a high-resolution image; The second segmentation module is used to input the high-resolution image into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask for the second type of follicles in the high-resolution image; wherein, the second type of follicles are follicles with a size smaller than the first threshold. The coordinate mapping and output module is used to map the pixel-level segmentation mask back to the coordinate space of the original ultrasound image, generating a final antral follicle segmentation map that corresponds one-to-one with the original ultrasound image in spatial location.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the super-resolution-based antral follicle identification method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the super-resolution-based antral follicle identification method as described in any one of claims 1 to 7.

Images