Gynecological tumor lymph node metastasis detection method based on edge feature extraction and fusion

Abstract

The invention provides a gynecological tumor lymph node metastasis detection method based on edge feature extraction and fusion, and belongs to the technical field of medical image detection technologies. And preprocessing the pelvic cavity enhanced CT image, and inputting the processed image into a pre-trained gynecological tumor pelvic lymph node metastasis accurate detection model. According to the accurate detection model for the pelvic lymph node metastasis of the gynecological tumor, multi-scale features and edge features of lymph nodes are extracted by using a multi-scale feature and edge information extraction convolution technology by virtue of a newly designed backbone network, and multi-scale feature fusion is carried out through a feature channel selection aggregation module. The features of the neck network fusion are transmitted to a detection head for analysis, the detection head deeply analyzes the features, and finally the detection result of the gynecological tumor lymph node metastasis is obtained. According to the pelvic lymph node detection method, the edge details and the spatial position of the pelvic lymph node can be accurately captured, and the detection precision is greatly improved through accurate analysis of the lymph node form and the edge features when an enhanced CT image of a patient is processed.

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging detection technology, relates to lymph node metastasis-related detection technology, and particularly relates to a gynecological tumor lymph node metastasis detection method based on edge feature extraction and fusion. Background Art

[0002] Gynecological cancers, primarily including cervical, ovarian, and endometrial cancers, are a major challenge to women's health worldwide. Lymph node metastasis is a common route of gynecological tumor spread, primarily affecting the pelvic lymph nodes, which are primarily distributed in the para-aortic, common iliac, internal and external iliac, and obturator lymph nodes. Adjacent tissues primarily include paravascular and pelvic connective tissue and adipose tissue. The diagnosis of lymph node metastasis primarily relies on postoperative pathological biopsy, but this method has a significant lag and cannot provide sufficient basis for preoperative treatment planning. Therefore, accurate preoperative prediction of pelvic lymph node metastasis is particularly important.

[0003] In past clinical practice, traditional imaging methods such as ultrasound, MRI, and CT were the primary means of detecting pelvic lymph node metastasis. Although ultrasound is easy to use, it is easily interfered with by gas and fatty tissue, making it difficult to clearly image deep lymph nodes. MRI images have relatively low spatial resolution, making it difficult to identify metastatic lymph nodes with blurred boundaries and small size. Although CT can provide three-dimensional images, the density difference between lymph nodes and surrounding fat and blood vessels is not obvious, which greatly increases the difficulty of detecting obese patients. In addition, these traditional detection methods are greatly affected by equipment parameters and the imaging environment. Noise and artifacts in the image reduce the accuracy of feature extraction, making it difficult to accurately detect tiny metastatic lesions.

[0004] To address the shortcomings of traditional detection methods, detection technology based on machine learning has emerged. Machine learning algorithms can automatically process the detection process, analyze massive amounts of data, and reduce interference from human factors. However, in practical applications, this technology faces many challenges. The internal structure of the pelvis is complex and the imaging information is rich. It is difficult for machine learning models to accurately extract the characteristics of micrometastases, and the detection effect is unsatisfactory. There are differences in equipment and scanning parameters between different hospitals. In addition, the individual physiological structures of patients are different. The generalization ability of the model is poor, and the reliability of the test results is difficult to guarantee. In addition, the operation process of some machine learning models is complex and the computational efficiency is low, which cannot meet the needs of rapid clinical diagnosis.

[0005] In recent years, deep learning technology has been widely used in the field of medical image analysis. Deep learning-based models, with their powerful feature extraction capabilities, have brought new solutions to the detection of pelvic lymph node metastases in gynecological tumors. However, existing deep learning models still face some challenges when detecting pelvic lymph node metastases in gynecological tumors. First, the contrast of pelvic lymph node metastases is generally low, and when processing such images, models are prone to insufficient feature extraction and inaccurate positioning, resulting in missed detection of small metastases. Second, due to the large variability in the morphology and size of pelvic lymph nodes across patients, existing models struggle to adapt to these variations, and their generalization capabilities need to be improved. Therefore, improving detection efficiency while maintaining detection accuracy has become a pressing issue in the field of pelvic lymph node metastasis detection in gynecological tumors. This project aims to develop an efficient detection method by optimizing feature extraction and fusion mechanisms, introducing a multi-scale attention mechanism, and designing a lightweight network structure. This approach will provide clinicians with accurate and rapid diagnostic tools and promote advancements in gynecological tumor diagnosis and treatment. Summary of the Invention

[0006] The purpose of the present invention is to solve the problems existing in the above-mentioned prior art and propose a gynecological tumor lymph node metastasis detection method based on edge feature extraction and fusion, so as to improve the detection accuracy of gynecological tumor lymph node metastasis detection.

[0007] In order to solve the above problems, the present invention implements the process as follows:

[0008] A method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion, comprising the following steps:

[0009] Step S1: Collect enhanced CT images of various gynecological cancer patients, annotate the lymph nodes in the enhanced CT images, and construct a benchmark dataset.

[0010] In step S2, the YOLO11 target detection model is used as the basic architecture to improve the backbone network, neck network, and loss function to build an accurate detection model for pelvic lymph node metastasis of gynecological tumors.

[0011] Step S3: Based on the benchmark data set constructed in step S1, the accurate detection model for pelvic lymph node metastasis of gynecological tumors obtained in step S2 is trained to obtain the final accurate detection model for pelvic lymph node metastasis of gynecological tumors.

[0012] In step S4, enhanced CT images of patients are collected to construct a test dataset, and a comprehensive evaluation and comparison are performed on the final accurate detection model for pelvic lymph node metastasis of gynecological tumors obtained in step S3, and the results are visualized and analyzed.

[0013] Furthermore, the step S1 is specifically as follows:

[0014] Enhanced CT images of patients with cervical, ovarian, or endometrial cancer were collected. The original enhanced CT images were in DICOM format and converted to a natural image format by adjusting the window width and window position. The pelvic region was then cropped and the image size was adjusted using bicubic interpolation. Finally, the pelvic lymph nodes were annotated using annotation tools to form a gynecological tumor lymph node detection image dataset, which was then used as a benchmark dataset.

[0015] In step S1, enhanced CT image data from a variety of gynecological cancer patients, including those at different clinical stages, is collected. A labeling tool is used to annotate the collected image data, selecting pelvic lymph nodes with rectangular boxes and noting whether the nodes are metastatic.

[0016] Furthermore, the step S2 is specifically as follows:

[0017] The proposed model for accurately detecting pelvic lymph node metastasis in gynecological tumors includes a backbone network, a neck network, and a detection head, connected in series. The first two modules of the backbone network are a multi-scale feature and edge information extraction convolutional module (MFEConv), which improves the backbone network using the multi-scale feature and edge information extraction convolutional module. The multi-scale feature and edge information extraction convolutional module includes two branches: a multi-scale feature extraction branch and an edge information extraction branch. The backbone and neck networks are connected via two feature channel selection aggregation modules (FCSMs). Finally, the regression loss function in the detection head is the NWD-CIoU loss.

[0018] By introducing a multi-scale feature and edge information extraction convolution module into the backbone network of the YOLO11 object detection model, a feature channel selection aggregation module into the neck network, and the normalized Gaussian Wasserstein distance (NWD) into the loss function, an NWD-CIoU regression loss function was constructed. The components of the gynecological tumor pelvic lymph node metastasis accurate detection model are as follows:

[0019] The backbone network is composed of two MFEConv modules connected in series, and then connected in series with a C3k2 module, a Conv module, a C3k2 module, a Conv module, a C3k2 module, an SPPF module and a C2PSA module.

[0020] Furthermore, the Conv module in the backbone network includes a labeled convolution layer, a batch normalization layer and an activation function SILU layer connected in series; the C3k2 module includes a Conv module and a C3K module; and the C2PSA module includes a Conv module and a PSA module connected in series.

[0021] Furthermore, the MFEConv module in the backbone network includes six parallel branches; first, the input feature map enters a 3×3 convolution layer with a stride of 2. This operation performs spatial downsampling, reducing the image resolution to half of the original, generating intermediate features, effectively reducing the amount of data, while retaining the key structural information of the image to prepare for subsequent feature extraction. Subsequently, the data is diverted into a carefully designed feature extraction module, namely three parallel differential convolution branches and one parallel multi-scale feature extraction convolution branch. The three differential convolution branches all use a 3×3 convolution kernel, and through clever weight configuration, they capture the horizontal gradient, vertical gradient, and angular gradient information of the image respectively. This design integrates the advantages of traditional edge detection operators based on gradient calculation. Compared with ordinary convolution, it can focus on the edge features of the lymph nodes more accurately when facing pelvic lymph node images. The parallel multi-scale feature extraction convolution branch consists of a channel-by-channel convolution with a kernel size of 3 and a channel-by-channel convolution with a kernel size of 5, connected in series. The third branch consists of a channel-by-channel convolution with a kernel size of 3, a channel-by-channel convolution with a kernel size of 5, and a channel-by-channel convolution with a kernel size of 7, connected in parallel to form the multi-scale feature extraction convolution branch. The image is analyzed at multiple scales. Finally, the differential convolution branch and the parallel multi-scale feature extraction convolution branch are element-wise added, and the output feature map is obtained through a normal convolution with a kernel size of 1. The output feature map is then sequentially passed through a series of C3k2 modules, Conv modules, C3k2 modules, Conv modules, C3k2 modules, SPPF modules, and C2PSA modules to extract deeper semantic information. The C2PSA module combines the CSP concept with the self-attention mechanism to enhance feature fusion, focus on key features, and improve the model's feature extraction and representation capabilities for the target.

[0022] Specifically, the multi-scale feature extraction branch can extract multi-scale features of enhanced CT images. By constructing a parallel multi-branch structure, each branch is configured with convolution kernels of different sizes to achieve multi-scale sampling of enhanced CT images. Smaller-sized convolution kernels have a relatively small receptive field during convolution operations and can calculate relatively localized pixel points in the enhanced CT image, thereby obtaining detailed features in the enhanced CT image. In contrast, larger-sized convolution kernels have a larger receptive field and can include pixel information from a wider area of the enhanced CT image in the calculation range during convolution operations, capturing the overall shape of the target in the enhanced CT image and its spatial relationship with the surrounding environment. During operation, each branch focuses on extracting local contextual information at a specific scale. This parallel and differentiated processing method overcomes the disadvantage of a single convolution kernel that can only obtain features at a single scale. Multiple convolution kernels of different scales working in parallel can deeply analyze the characteristics of the target lymph node area in the enhanced CT image from multiple dimensions such as space and frequency, without missing tiny details while fully grasping the overall structure, greatly improving the comprehensiveness and depth of feature extraction from complex enhanced CT images.

[0023] Specifically, the edge information extraction branch in the backbone network is composed of three parallel differential convolutions. The differential convolution is based on the unique design of the convolution operation and can keenly capture subtle changes in the edges of lymph node targets in enhanced CT images. When faced with images of pelvic lymph node metastases, there are extremely subtle differences between the edges of lesions and normal lymph nodes at the transition between different grayscales and textures. The differential convolution, with its high sensitivity to gradient changes, converts these changes into deep edge semantic features. By setting up multiple differential convolutions in parallel, edge information of different scales can be extracted simultaneously, from the macroscopic lymph node contour to the microscopic metastasis boundary, and the multi-dimensional edge features complement each other.

[0024] Specifically, the calculation formulas for differential convolution and ordinary convolution are as follows:

[0025]

[0026] Among them, ω i is the value of the convolution kernel, x i is the pixel value in the input feature map, x i -x i ′ is the pixel difference in the specified direction.

[0027] Specifically, the calculation formula of the MFEConv module is as follows:

[0028] X m =Conv 3*3 (X input )#

[0029] Xdf =VDC(X m )+HDC(X m )+ADC(X m )#

[0030] X p =DWC 7*7 (DWC 5*5 (DWC 3*3 (X m )))+DWC 5*5 (DWC 3*3 (X m ))+DWC 3*3 (X m )#

[0031] X output =Conv 1*1 (PWC(X p )+X df )

[0032] Among them, X input Represents the input feature map, Conv 3*3 and Conv 1*1 represents a common convolution layer with kernel sizes of 3 and 1 respectively, X m Represents the output feature map after ordinary convolution, VDC, HDC and ADC represent vertical differential convolution, horizontal differential convolution and center differential convolution respectively, X df Represents the output feature map of the edge information extraction branch, DWC 3*3 、DWC 5*5 and DWC 7*7 They represent channel-by-channel convolution with kernel sizes of 3, 5, and 7, respectively. p Represents the output feature map of the multi-scale feature extraction branch, X output Represents the output feature map.

[0033] Furthermore, the overall structure of the neck network first uses Feature Pyramid Networks to fuse feature maps through a bottom-up path. 80×80 low-level features are gradually transferred to 40×40 and 20×20 high-level features, forming a preliminary feature pyramid. This pyramid enables the model to utilize information from different scales to recognize objects. A Path Aggregation Network is then used for top-down feature transfer and fusion, gradually transferring 20×20 high-level semantic features to 40×40 and 80×80 low-level features. This low-level features not only contain their own detailed information but also incorporate rich semantic information from higher levels. This top-down path aggregation enables low-level feature maps to better capture information from small objects and complex backgrounds while maintaining the global perspective of high-level semantic features. The Path Aggregation Network structure fuses multiple levels of features through a top-down path, enabling the model to handle both large and small objects. Generate feature maps of different scales, such as 80×80, 40×40, and 20×20.

[0034] The neck network is composed of an Upsample module, an FCSM module, a C3k2 module, an Upsample module, an FCSM module, a C3k2 module, a Conv module, a Concat module, a C3k2 module, a Conv module, a Concat module and a C3k2 module connected in series.

[0035] The detailed structure of the FCSM module in the neck network is as follows: First, high-dimensional feature maps and low-dimensional feature maps are processed differently. High-dimensional feature maps are divided into four groups based on their channel dimension. High-dimensional feature maps often contain more abstract and richer semantic information about the image, and this grouping facilitates subsequent refined processing of features specific to different parts. As for low-dimensional feature maps, since their channel count may not match that of high-dimensional feature maps, they are first adjusted using a standard convolutional module to better integrate them with the high-dimensional feature maps in subsequent processing. After this adjustment, they are also divided into four groups.

[0036] After grouping is complete, the feature fusion stage begins. For each feature group, the high-dimensional feature map is element-wise added to the low-dimensional feature map. This addition allows the high- and low-dimensional feature maps to complement each other, combining the semantic information of the high-dimensional feature map with the detailed information of the low-dimensional feature map. Subsequently, dynamic channel weights are independently generated for each feature group through the channel attention mechanism SE (Squeeze-and-Excitation). The channel attention mechanism dynamically measures the importance of each channel based on the feature distribution of the feature group itself. Different feature groups may have different importance distributions in different task scenarios. By dynamically adjusting the channel weights, the fusion ratio of high and low feature maps can be dynamically adjusted based on their own feature distribution, allowing the gynecological tumor pelvic lymph node metastasis accurate detection model to more intelligently perform feature fusion. To ensure the stability and accuracy of the training of the gynecological tumor pelvic lymph node metastasis accurate detection model, the FCSM module adds residual connections to the fusion of high- and low-dimensional feature maps. During the training of deep learning models, the vanishing gradient problem is prone to occur during backpropagation, which can lead to model convergence difficulties or poor training results. Residual connections enable gradients to flow more smoothly during backpropagation, effectively avoiding the vanishing gradient problem, ensuring that the accurate detection model for pelvic lymph node metastasis of gynecological tumors can be trained stably, and improving the performance of the accurate detection model for pelvic lymph node metastasis of gynecological tumors.

[0037] Finally, each set of processed outputs is concatenated along the channel dimension, reintegrating the features from different groups to produce the final output. This output combines the advantages of both high- and low-dimensional features, and undergoes intelligent weighting adjustments, providing higher-quality feature information for subsequent image analysis tasks.

[0038] Specifically, the calculation formula of the FCSM module is as follows:

[0039] F high =[H1,H2,H3,H4]#

[0040] F low =[L1,L2,L3,L4]=Conv 3*3 (F input-low )#

[0041] ω=SE(H i +L i )#

[0042] F output =Concat(ω*H i +(1-ω)*L i +H i +L i )#

[0043] Among them, [] means grouping the feature maps by channels, F high Represents the high-dimensional feature map after grouping, F input-low Represents the initial low-dimensional feature map, Conv 3*3 represents a normal convolution layer with a convolution kernel of 3, F low Represents the low-dimensional feature map after grouping, H i and L i They represent a group of high-dimensional feature maps and low-dimensional feature maps after grouping, SE represents the channel attention module, ω represents the weight generated by the channel attention module, and Concat represents the concatenation of multiple feature maps by channel. output Represents the output feature map of the FCSM module.

[0044] The regression loss function uses NWD-CIoU as the regression loss function. The CIoU loss function reveals significant flaws. Due to the significant variation in pelvic lymph node morphology and size, and the presence of numerous tiny lymph nodes, the CIoU loss function is extremely sensitive to small target position deviations. When calculating the position deviation between the predicted and true frames of small lymph nodes, the gradient calculation experiences drastic fluctuations. During the backpropagation phase of training the gynecological tumor pelvic lymph node metastasis precision detection model, this unstable gradient calculation severely hinders the effective updating of the model's parameters, significantly reducing the model's detection accuracy for small lymph nodes and thus affecting overall detection performance.

[0045] Furthermore, to overcome the above difficulties, the present invention designs an NWD-CIoU composite regression loss function, and innovatively introduces the normalized Gaussian Wasserstein distance (NWD) into the loss function. This measure avoids gradient anomalies caused by small target position deviations by more accurately measuring the distance between the predicted box and the true box; the designed dual-weighted loss function structure can dynamically adjust the attention of the gynecological tumor pelvic lymph node metastasis accurate detection model to changes in the positions of targets of different sizes based on the characteristics of small targets and conventional targets.

[0046] Specifically, the NWD-CIoU composite regression loss function is calculated as follows:

[0047]

[0048] Among them, cx a 、cy a 、w a 、h a are the center horizontal coordinate, center horizontal coordinate, width and height of the real frame, cx b 、cy b 、w b 、hb are the center horizontal coordinate, center horizontal coordinate, width and height of the prediction box respectively, and is the real box and the predicted box, W2 2 is the second-order Wasserstein distance between the real box and the predicted box, C is a constant, is the normalized second-order Wasserstein distance, is the NWD loss in the regression loss, α NWD It is a parameter to adjust CIoU loss and NWD loss, with a value of 0.5. is the NWD-CIoU regression loss.

[0049] The effects and benefits of the present invention are:

[0050] (1) The present invention provides a precise detection model for pelvic lymph node metastasis of gynecological tumors based on edge feature extraction and fusion. A multi-scale feature and edge information extraction convolution (MFEConv) module is designed. In enhanced CT images, pelvic lymph nodes vary in shape and size, and the edge features of micro-metastatic lesions are extremely critical. The MFEConv module can perform multi-scale sampling on the image by connecting multiple depth-wise convolutions of different scales in parallel. Smaller-scale depth-wise convolutions can keenly capture the subtle texture and edge details of the lymph nodes; larger-scale depth-wise convolutions can grasp the overall shape and spatial position of the lymph nodes. At the same time, multiple differential convolution modules operate in parallel. Differential convolutions are extremely sensitive to the subtle differences between metastatic and non-metastatic lymph nodes, and can extract subtle differences between the two in terms of edge texture, grayscale transition, etc. This design greatly improves the backbone network's ability to extract features from CT images.

[0051] (2) The FCSM module is designed in the accurate detection model of pelvic lymph node metastasis of gynecological tumors designed by the present invention. In the process of multi-scale feature fusion, feature maps of different scales contain information at different levels, which is crucial to the performance of the model. The FCSM module uses the adaptive channel weight control mechanism to automatically generate dynamic weights for the feature distribution of each channel. When fusing multi-scale features, the module intelligently selects key information based on these weights, suppresses redundant information, and more accurately retains important discriminative features. When processing medical images containing lymph node metastases, the FCSM module can strengthen the feature channels related to the metastases and weaken irrelevant channels, significantly improving the quality of feature fusion, effectively improving the accuracy of the model in predicting pelvic lymph node metastasis, and avoiding misjudgment due to feature confusion.

[0052] (3) The accurate detection model of pelvic lymph node metastasis of gynecological tumors designed in this paper constructs the NWD-CIoU composite loss function; the NWD-CIoU composite loss function introduces the normalized Gaussian Wasserstein distance (NWD), which can more accurately measure the distance between the predicted box and the true box, and effectively compensate for the position sensitivity defect of traditional IoU to small targets; when detecting small metastatic lesions of pelvic lymph nodes, the NWD-CIoU composite loss function can make the gradient calculation in the model training process more stable, optimize the model parameter update, and significantly enhance the positioning accuracy of the lymph node area, thereby achieving accurate prediction of pelvic lymph node metastasis.

[0053] (4) The precise detection model for pelvic lymph node metastasis of gynecological tumors designed by the present invention has achieved optimized performance in many aspects. The synergistic effect of the MFEConv module, the FCSM module and the NWD-CIoU composite loss function has improved the accuracy of feature extraction, feature fusion and positioning. While achieving accurate prediction of pelvic lymph node metastasis, the operating efficiency of the precise detection model for pelvic lymph node metastasis of gynecological tumors has been improved; the precise detection model for pelvic lymph node metastasis of gynecological tumors can not only quickly process a large number of pelvic enhanced CT images, provide doctors with accurate lymph node metastasis prediction results, help doctors formulate more scientific treatment plans, but also reduce the requirements for hardware equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is an overall flow chart of an embodiment of the present invention;

[0055] Figure 2 The data processing operation flow of an embodiment of the present invention;

[0056] Figure 3 Schematic diagram of the overall structure of a method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to an embodiment of the present invention;

[0057] Figure 4 This is the differential convolution calculation process of an embodiment of the present invention;

[0058] Figure 5 Schematic diagram of the structure of the MFEConv module according to an embodiment of the present invention;

[0059] Figure 6 Schematic diagram of the structure of the FCSM module according to an embodiment of the present invention;

[0060] Figure 7 Schematic diagram of the structure of the target detection head module according to an embodiment of the present invention;

[0061] Figure 8 This is a schematic diagram for visualizing the detection results of an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0063] In the development of gynecological tumors, pelvic lymph node metastasis is a key factor influencing clinical staging, treatment decisions, and patient prognosis. Pelvic lymph node metastasis often indicates that the disease has entered a more advanced stage. This not only significantly increases the difficulty of surgical resection and places higher demands on the formulation of chemoradiotherapy, but also greatly increases the patient's risk of mortality. Therefore, accurately predicting pelvic lymph node metastasis is extremely important for improving patient prognosis.

[0064] The YOLO (You Only Look Once) model is a target detection model with the advantages of strong real-time performance, high accuracy, and lightweight. It has been used in a wide range of application scenarios, including object recognition, human body detection, and lesion identification. The YOLO model provides a more advanced solution for the detection of pelvic lymph node metastases in gynecological tumors. Compared with traditional lymph node metastasis detection methods, it is more intelligent and efficient, and can quickly formulate personalized treatment plans, buying valuable time for patients' follow-up treatment and improving their prognosis. However, pelvic lymph nodes themselves are relatively small, and metastatic lesions are often closely adhered to surrounding tissues such as fat and blood vessels. This causes an imbalance in the contrast between the lesions and surrounding tissues in the image, obscures a large amount of critical edge and texture information, and greatly increases the difficulty of detection, posing a great challenge to clinical diagnosis.

[0065] Based on the above considerations, the present invention provides a method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion. The specific implementation process is as follows:

[0066] This invention closely revolves around the actual needs of clinical diagnosis of gynecological tumors, combines the diagnosis and treatment environment and data characteristics of the hospital, and innovatively proposes a detection algorithm for the field of gynecological tumor pelvic lymph node metastasis detection - Gynecological Tumor Pelvic Lymph Node Metastasis Accurate Detection Model (PLNM-YOLO). This algorithm aims to break through the bottleneck of existing detection technology in terms of accuracy, and rely on the hospital's conventional computing power equipment to achieve accurate detection of gynecological tumor pelvic lymph node metastasis. PLNM-YOLO uses YOLO11 as the benchmark model and carries out systematic improvements on it. The overall architecture of PLNM-YOLO is as follows: Figure 3As shown. In the backbone network, PLNM-YOLO introduces channel-by-channel convolution and differential convolution modules. Differential convolution can keenly capture the subtle differences in the edges of metastatic and non-metastatic lymph nodes, further highlighting the edge features of the lymph nodes; channel-by-channel convolution can comprehensively and deeply mine the detailed features of the target area in the image from multiple dimensions. Based on this, the convolution module of the original backbone network is optimized, and a new multi-scale feature and edge information extraction module - Multi-scale Feature andEdge information extraction Convolution (MFEConv) is proposed. The detailed structure of MFEConv is shown in Figure 5 As shown in the figure, MFEConv replaces the first two ordinary convolution modules in YOLO11 to build a new backbone network. This design significantly improves the backbone network's ability to extract subtle changes in the edges of pelvic lymph nodes, especially micrometastases. In terms of the neck network, PLNM-YOLO designed a new feature fusion module - Feature Channel Selection Aggregation Module (FCSM). The detailed structure of FCSM is shown in the figure. Figure 6 As shown in , this module can dynamically adjust the fusion ratio of feature maps of different scales, so that PLNM-YOLO can capture the effective edge features of lymph nodes in all directions. In the detection head part, the NWD-CIoU composite regression loss function is designed. The detailed structure of the detection head is shown in Figure 7 As shown in the figure, small lymph node metastases are extremely small, and the traditional IoU loss function is extremely sensitive to small target position deviations, resulting in unstable gradient calculation during positioning. The NWD-CIoU composite loss function introduces the normalized Gaussian Wasserstein distance (NWD), which can more accurately measure the distance between the predicted and true bounding boxes. This effectively compensates for the shortcomings of the traditional IoU loss, making PLNM-YOLO more accurate during training and significantly improving the localization accuracy of lymph node regions, especially small metastases.

[0067] In an embodiment of the present invention, the role of the backbone network is to extract the features of the input image, and the quality of feature extraction affects the accuracy of subsequent network detection. The backbone network consists of two MFEConv modules connected in series, which are then connected in series with a C3k2 module, a Conv module, a C3k2 module, a Conv module, a C3k2 module, an SPPF module, and a C2PSA module. For the two MFEConv modules, the input feature map first enters a 3×3 convolution layer with a stride of 2. This operation performs spatial downsampling, reduces the image resolution to half of the original, generates intermediate features, effectively reduces the amount of data, and retains the key structural information of the image in preparation for subsequent feature extraction. Subsequently, the data is diverted into a carefully designed feature extraction module, namely three parallel differential convolution branches and one parallel multi-scale feature extraction convolution branch. Each of the three differential convolution branches uses a 3×3 convolution kernel. Through clever weight configuration, they capture the image's horizontal, vertical, and angular gradient information, respectively. This design integrates the advantages of traditional edge detection operators based on gradient calculation. Compared to ordinary convolution, it can more accurately focus on the edge features of lymph nodes when faced with pelvic lymph node images. The parallel multi-scale feature extraction convolution branch consists of a channel-by-channel convolution with a convolution kernel size of 3 and a channel-by-channel convolution with a convolution kernel size of 5 in series. The third branch is composed of a channel-by-channel convolution with a convolution kernel size of 3, a channel-by-channel convolution with a convolution kernel size of 5, and a channel-by-channel convolution with a convolution kernel size of 7 in series. These three branches are connected in parallel to form the multi-scale feature extraction convolution branch, which analyzes images at multiple scales. Finally, the differential convolution branch and the parallel multi-scale feature extraction convolution branch are element-wise added, and the output feature map is obtained by ordinary convolution with a convolution kernel size of 1; the output feature map is then sequentially passed through the series of C3k2 module, Conv module, C3k2 module, Conv module, C3k2 module, SPPF module and C2PSA module to extract deeper semantic information. The C2PSA module combines the CSP idea with the self-attention mechanism to enhance feature fusion, focus on key features, and improve PLNM-YOLO's feature extraction and expression capabilities for the target.

[0068] The advantages of the MFEConv structure designed in this invention over ordinary convolution are as follows:

[0069] (1) Design a multi-scale feature extraction branch. The multi-scale feature extraction branch consists of three branches. The first branch is a channel-by-channel convolution with a convolution kernel size of 3; the second branch is a channel-by-channel convolution with a convolution kernel size of 3 and a channel-by-channel convolution with a convolution kernel size of 5 in series; the third branch is a channel-by-channel convolution with a convolution kernel size of 3, a channel-by-channel convolution with a convolution kernel size of 5, and a channel-by-channel convolution with a convolution kernel size of 7 in series. The three branches are connected in parallel to form a multi-scale feature extraction branch. Convolution kernels of different sizes can perform multi-scale sampling on the image, and each branch focuses on extracting local context information at a specific scale. Compared with a single convolution kernel, this parallel and differentiated processing method can comprehensively and deeply mine the detailed features of the target area in the image from multiple dimensions.

[0070] (2) Design the edge information extraction branch. The edge information extraction branch consists of three parallel differential convolution modules: horizontal differential convolution with a convolution kernel of 3, vertical differential convolution, and angular differential convolution. Differential convolution can keenly capture subtle changes at the edges of objects in the image and convert these changes into deep edge semantic features.

[0071] In this embodiment of the present invention, the calculation formulas for differential convolution and ordinary convolution are as follows:

[0072]

[0073] where ω i is the value of the convolution kernel, x i is the pixel value in the input feature map, x i -x i ′ is the pixel difference in the specified direction. The calculation process of differential convolution is as follows Figure 4 shown.

[0074] In this embodiment of the present invention, the calculation formula of the MFEConv module is as follows:

[0075] X m =Conv 3*3 (X input )#

[0076] X df =VDC(X m )+HDC(X m )+ADC(X m )#

[0077] X p =DWC 7*7 (DWC 5*5 (DWC 3*3 (X m )))+DWC 5*5 (DWC 3*3(X m ))+DWC 3*3 (X m )#

[0078] X output =Conv 1*1 (PWC(X p )+X df )

[0079] Among them, X input Represents the input feature map, Conv 3*3 and Conv 1*1 represents a common convolution layer with kernel sizes of 3 and 1 respectively, X m Represents the output feature map after ordinary convolution, VDC, HDC and ADC represent vertical differential convolution, horizontal differential convolution and center differential convolution respectively, X df Represents the output feature map of the edge information extraction branch, DWC 3*3 、DWC 5*5 and DWC 7*7 They represent channel-by-channel convolution with kernel sizes of 3, 5, and 7, respectively. p Represents the output feature map of the multi-scale feature extraction branch, X output Represents the output feature map.

[0080] In the embodiment of the present invention, the role of the neck network is to perform feature fusion. The effect of feature fusion directly affects the ability of the accurate detection model for pelvic lymph node metastasis of gynecological tumors to comprehensively utilize features of different scales and the final detection performance.

[0081] In an embodiment of the present invention, the overall structure of the neck network first uses Feature Pyramid Networks to fuse feature maps through a bottom-up path. The 80×80 low-level features are gradually transferred to the 40×40 and 20×20 high-level features, forming a preliminary feature pyramid. This pyramid enables the model to utilize information from different scales to identify targets. Then, a Path Aggregation Network is used to perform top-down feature transfer and fusion, gradually transferring the 20×20 high-level semantic features to the 40×40 and 80×80 low-level features. In this way, the low-level features not only contain their own detailed information but also integrate rich semantic information from the high-level features. This top-down path aggregation enables the low-level feature maps to better capture information in small objects and complex backgrounds, while maintaining the global perspective of the high-level semantic features. The PAN structure fuses multi-level features through a top-down path, enabling PLNM-YOLO to process both large and small objects. Generate feature maps of different scales, such as 80×80, 40×40, and 20×20.

[0082] The detailed structure of the neck network sequentially passes through the Upsample module, FCSM module, C3k2 module, Upsample module, FCSM module, C3k2 module, Conv module, Concat module, C3k2 module, Conv module, Concat module, and C3k2 module. For the FCSM module, the high-dimensional feature map is first divided into four groups by channel. The low-dimensional feature map is also divided into four groups after adjusting the number of channels through a standard convolution module. The purpose of this grouping is to enable subsequent refined processing of different feature subsets, so that high- and low-dimensional features can better leverage their respective strengths during the fusion process. After grouping, for each feature group, the high-dimensional features are added element-by-element to the low-dimensional features. This addition operation allows the high- and low-dimensional features to complement each other, combining the abstract semantic information of the high-dimensional features with the detailed information of the low-dimensional features. Subsequently, dynamic channel weights are independently generated for each feature group using the channel attention mechanism (SE). The channel attention mechanism adaptively measures the importance of each channel based on the distribution of each feature group. This allows different groups to dynamically adjust the fusion ratio of high and low features based on their own circumstances, thereby more accurately retaining important discriminative features. Furthermore, the addition of residual connections during this process effectively prevents the vanishing gradient problem during backpropagation, ensuring the stability and accuracy of PLNM-YOLO training. Finally, the processed outputs of each group are concatenated along the channel dimension to form the final fused feature map.

[0083] The FCSM module designed in this invention has the following advantages over the Concat connection:

[0084] In the current feature fusion architecture, a simple channel splicing method is often used. This method only directly splices feature maps of different levels in the channel dimension. It lacks effective distinction and integration of feature importance, cannot fully explore the intrinsic connection between features at each level, and is difficult to achieve deep fusion of multi-level features. FCSM first reasonably groups high- and low-dimensional feature maps, and then adds elements one by one to allow features at different levels to complement each other. At the same time, with the help of the channel attention mechanism, the fusion ratio is dynamically adjusted according to the distribution characteristics of the features, effectively suppressing redundant information, accurately retaining key features, and improving the quality of feature fusion. In addition, FCSM adds residual connections, a design that can effectively prevent the problem of gradient disappearance during backpropagation, ensuring the stability and accuracy of model training.

[0085] In the embodiment of the present invention, the calculation formula of the FCSM module is as follows:

[0086] F high =[H1,H2,H3,H4]#

[0087] F low =[L1,L2,L3,L4]=Conv 3*3 (F input-low )#

[0088] ω=SE(H i +L i )#

[0089] F output =Concat(ω*H i +(1-ω)*L i +H i +L i )#

[0090] Among them, [] means grouping the feature maps by channels, F high Represents the high-dimensional feature map after grouping, F input-low Represents the initial low-dimensional feature map, Conv 3*3 represents a normal convolution layer with a convolution kernel of 3, F low Represents the low-dimensional feature map after grouping, H i and L i They represent a group of high-dimensional feature maps and low-dimensional feature maps after grouping, SE represents the channel attention module, ω represents the weight generated by the channel attention module, and Concat represents the concatenation of multiple feature maps by channel. output Represents the output feature map of the FCSM module.

[0091] In an embodiment of the present invention, the regression loss function in the loss function is a NWD-CIoU composite regression loss function. The traditional CIoU loss function is very sensitive to this type of small target position deviation. When calculating the position deviation between the small target prediction box and the real box, a small position change will cause a sharp fluctuation in the gradient, resulting in unstable gradient calculation. During the model backpropagation process, this unstable gradient makes it impossible to effectively update the model parameters, reduce the model's detection accuracy for small targets, and affect the model's accurate judgment of lymph node metastasis. Therefore, the present invention designs an NWD-CIoU composite regression loss function and introduces the normalized Wasserstein distance into the loss function. This method can more accurately measure the distance between the prediction box and the real box, and effectively avoid gradient anomalies caused by small target position deviations. At the same time, a dual-branch weight structure is designed, and the NWD branch and the CIOU branch focus on the spatial positioning accuracy and geometric adaptation of the detection box respectively.

[0092] In this embodiment of the present invention, the calculation formula of the NWD-CIoU composite regression loss function is as follows:

[0093]

[0094] where cx a 、cy a 、w a 、h a are the center horizontal coordinate, center horizontal coordinate, width and height of the real frame, cx b 、cy b 、w b 、h b are the center horizontal coordinate, center horizontal coordinate, width and height of the prediction box respectively, and is the real box and the predicted box, W2 2 is the second-order Wasserstein distance between the real box and the predicted box, C is a constant, is the normalized second-order Wasserstein distance, is the NWD loss in the regression loss, α NWD It is a parameter to adjust CIoU loss and NWD loss, with a value of 0.5. is the NWD-CIoU regression loss.

[0095] In this embodiment of the present invention, a detailed gynecological tumor lymph node metastasis detection process is described. First, a set of anchor boxes are generated on feature maps at different scales to predict the target's position and size. Bounding box regression is performed on the anchor boxes using the NWD-CIoU loss function to obtain the target bounding box, which is a prediction of the relative position and size. The predicted coordinates include the center point (x, y), width (W), and height (H), representing the position and size of the detection box.

[0096] Through the classification path in the detection head, each or anchor box is mapped to a category probability distribution, the Softmax function (normalized exponential function) is used to convert the output score of the model into category probability, and the cross entropy loss is used to measure the difference between the probability distribution of the predicted category and the true category distribution.

[0097] For each anchor box, a confidence value needs to be output. This value is normalized to [0, 1] through the Sigmoid function (S-shaped growth curve function), which represents the probability of the existence of the target in the box. In this embodiment of the present invention, it is the probability of the existence of a lymph node.

[0098] The prediction results are then filtered through non-maximum suppression (NMS), and the predicted bounding boxes are NMS-ed to remove the boxes with large overlaps and only retain the boxes with the highest confidence to ensure that the final output results are accurate and non-redundant.

[0099] In the embodiment of the present invention, the experiment was conducted in a software and hardware environment with an Intel Core i9-13900K CPU, an NVIDIA RTX4090 (24GB) GPU, CUDA11.8, Python3.9, and an Ubuntu22.04.1 operating system. The training set was input into the improved network model for forward propagation training, and the improved network model was updated and optimized using the backpropagation algorithm; and the improved network model was iteratively trained multiple times using the NWD-CIoU loss function and the SGD optimizer until the improved network model converged to obtain a trained optimal weight file. In the embodiment of the present invention, the PLNM-YOLO model designed by the present invention was used as a training model, the initial learning rate was set to 0.01, the batch size was set to 16, the number of iterations was 400, the IOU threshold was 0.7, and the optimal network model weight file was obtained through multiple rounds of iterative training, thereby obtaining the optimal gynecological tumor lymph node metastasis detection model.

[0100] Experimental results verify:

[0101] In an embodiment of the present invention, the performance evaluation indicators include a performance evaluation of the target network model after the test through preset performance evaluation indicators, wherein the performance evaluation indicators include accuracy P (Precision), recall R (Recall), F1 (F1_score) score, mean average precision mAP (mean Average Precision), and model inference speed (FPS).

[0102] The definition of accuracy is:

[0103] Recall is defined as:

[0104] The F1 score (F1_score) is defined as:

[0105] The definition of average precision AP is:

[0106] The mean average precision (mAP) is defined as: Where TP is the true positive, that is, the positive sample is correctly identified as a positive sample; FN is the false negative, that is, the positive sample is mistakenly identified as a negative sample; TN is the true negative, that is, the negative sample is correctly identified as a negative sample; FP is the false positive, that is, the negative sample is mistakenly identified as a positive sample.

[0107] To further analyze the effectiveness of each module, we conducted ablation experiments on each module to verify whether it effectively functions within the model. We calculated the corresponding performance metrics for each model, retaining different modules. The ablation results are shown in Table 1.

[0108] Table 1 Ablation experiment

[0109] Model YOLO11 MFEConv FCSM NWD P R mAP@50 A √ 0.743 0.653 0.733 B √ √ 0.819 0.605 0.756 C √ √ √ 0.818 0.659 0.777 D √ √ √ √ 0.828 0.684 0.799

[0110] Ablation experiments show that the model with the multi-scale feature and edge information extraction convolution module improves mAP@50 by 2.3%. Furthermore, the introduction of the feature channel selection aggregation module further improves mAP@50 by 4.4%. Finally, the NWD-CIoU loss is introduced to obtain the PLNM-YOLO model designed by this invention, which further improves mAP@50 by 6.6%.

[0111] To systematically and comprehensively test the performance improvements achieved by the method of the present invention in detecting gynecological tumor lymph node metastasis, key performance indicators such as accuracy and recall were calculated for both current mainstream detection models and the innovative PLNM-YOLO model developed in the present invention. Table 2 clearly shows the comparison of these indicators between existing technologies and the method of the present invention.

[0112] Table 2 Comparative experiment

[0113] method P R F1 mAp@50 FPS Faster R-CNN 0.587 0.542 0.564 0.568 206 Retinanet 0.502 0.486 0.494 0.463 313 Fcos 0.566 0.537 0.551 0.550 268 RT-DETR 0.732 0.692 0.711 0.715 613 YOLOv8n 0.726 0.665 0.694 0.717 586 YOLO10n 0.763 0.659 0.709 0.699 601 PLNM-YOLO 0.828 0.684 0.749 0.799 652

[0114] As shown in Table 2, the PLNM-YOLO model in this embodiment of the present invention achieves higher accuracy in detecting and identifying gynecological tumor lymph node metastasis than the FasterR-CNN algorithm. It also achieves higher accuracy than other YOLO algorithms, enabling more accurate detection and location of bubbles. The PLNM-YOLO model also boasts faster detection speeds and is easier to deploy and apply. This model is more feasible and superior to existing technologies.

[0115] As can be seen from the above examples, the present invention discloses a method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion, including: obtaining a CT image dataset of gynecological tumor patients; replacing the first two ordinary convolutional layers in the original model backbone network structure with the MFEConv module designed by the present invention to effectively extract multi-scale features and edge information; using the Concat feature splicing layer in the neck network structure of the original model of the FCSM module designed by the present invention to enable the model to dynamically fuse high-level semantic features and low-level detail features; and using NWD to improve the loss function in the model, designing a new NWD-CIoU composite loss function to enhance the model's adaptability to changes in the location of gynecological tumor lymph node detection. The training set is input into the improved network for training, and the test set is input into the trained network for performance evaluation. The present invention achieves a gynecological tumor lymph node metastasis detection effect with high computational accuracy and high speed under the same experimental software and hardware platforms.

[0116] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion, characterized in that: The following steps are involved: Step S1: Collect enhanced CT images of various gynecological cancer patients, annotate the lymph nodes in the enhanced CT images, and construct a benchmark dataset; Step S2: Using the YOLO11 target detection model as the basic architecture, the backbone network, neck network, and loss function are improved to build an accurate detection model for pelvic lymph node metastasis of gynecological tumors; Step S3, based on the benchmark data set constructed in step S1, the accurate detection model for pelvic lymph node metastasis of gynecological tumors obtained in step S2 is trained to obtain the final accurate detection model for pelvic lymph node metastasis of gynecological tumors; In step S4, enhanced CT images of patients are collected to construct a test dataset, and a comprehensive evaluation and comparison are performed on the final accurate detection model for pelvic lymph node metastasis of gynecological tumors obtained in step S3, and the results are visualized and analyzed.

2. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 1, characterized in that: In the step S1 , enhanced CT image data of a variety of different gynecological cancer patients are used, including enhanced CT image data of patients at different clinical stages.

3. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 1, characterized in that: In step S1 , when collecting image data of patients with various gynecological cancers and various clinical stages, a labeling tool is used to label the collected image data, and a pelvic lymph node is selected by a rectangular frame and marked whether the lymph node has metastasized.

4. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 1, characterized in that: In step S2, the accurate detection model for pelvic lymph node metastasis of gynecological tumors includes a backbone network, a neck network, and a detection head connected in series, wherein the backbone network and the neck network are connected via a feature channel selection aggregation model (FCSM); the improvement is: The multi-scale feature and edge information extraction convolution module MFEConv is introduced into the backbone network of the YOLO11 target detection model. The multi-scale feature and edge information extraction convolution module includes two branches: a multi-scale feature extraction branch and an edge information extraction branch. Introducing the feature channel selection aggregation module FCSM in the neck network; The normalized Gaussian Wasserstein distance (NWD) is introduced into the loss function of the detection head, and the NWD-CIoU regression loss function is constructed to perform refined regression on the pelvic lymph node detection box.

5. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 4, characterized in that: The backbone network is composed of a multi-scale feature and edge information extraction convolution module connected in series, and then connected in series in sequence to form a C3k2 module, a Conv module, a C3k2 module, a Conv module, a C3k2 module, an SPPF module, and a C2PSA module; The multi-scale feature extraction branch consists of three parallel depth-wise convolutions, and the convolution kernels of the three parallel depth-wise convolutions are 3, 5, and 7 respectively; the edge information extraction branch consists of parallel vertical differential convolution, horizontal differential convolution, and angular differential convolution, and the convolution kernel size of the three differential convolutions is 3, which is used to extract multi-scale features of the enhanced CT image; By constructing a parallel multi-branch structure, each branch is configured with a convolution kernel of different sizes to achieve multi-scale sampling of enhanced CT images; The edge information extraction branch is composed of three parallel differential convolutions and is used to extract subtle edge features of pelvic lymph nodes.

6. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 4, characterized in that: The calculation formula of the edge information extraction convolution module MFEConv is as follows: X m =Conv 3*3 (X input )# X df =VDC(X m )+HDC(X m )+ADC(X m )# X p =DWC 7*7 (DWC 5*5 (DWC 3*3 (X m )))+DWC 5*5 (DWC 3*3 (X m )) + DWC 3*3 (X m )# X output =Conv 1*1 (PWC(X p )+X df ) Among them, X input Represents the input feature map, Conv 3*3 and Conv 1*1 represents a common convolution layer with kernel sizes of 3 and 1, respectively, X m Represents the output feature map after ordinary convolution, VDC, HDC and ADC represent vertical differential convolution, horizontal differential convolution and center differential convolution respectively, X df Represents the output feature map of the edge information extraction branch, DWC 3*3 、DWC 5*5 and DWC 7*7 They represent channel-by-channel convolution with kernel sizes of 3, 5, and 7, respectively. p Represents the output feature map of the multi-scale feature extraction branch, X output Represents the output feature map.

7. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 4, characterized in that: The feature channel selection aggregation module FCSM in the neck network divides high-dimensional features and low-dimensional features into four groups, each of which is weightedly fused using channel attention and then fused using residual connections. Specifically: First, high-dimensional feature maps and low-dimensional feature maps are processed differently. For high-dimensional feature maps, they are divided into four groups according to the channel dimension. For low-dimensional feature maps, the number of channels is adjusted through the ordinary convolution module, and then they are also divided into four groups after adjustment. After the grouping is completed, the feature fusion stage begins; for each group of features, the high-dimensional feature map and the low-dimensional feature map are added element by element; Subsequently, dynamic channel weights are independently generated for each set of features through the channel attention mechanism (SE). To ensure the stability and accuracy of the training of the precise detection model for pelvic lymph node metastasis in gynecological tumors, the FCSM module adds residual connections to the fusion of high-dimensional feature maps and low-dimensional feature maps. Finally, each group of processed outputs is spliced in the channel dimension, and the features of different groups are reintegrated together to obtain the final output result.

8. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 7, characterized in that: The calculation formula of the FCSM module is as follows: F high =[H1,H2,H3,H4]# <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> low <h2 style=";text-align:left;direction:ltr"> ([L1,L2,L3,L4]) = Conv<h2 style=";text-align:left;direction:ltr"> 3*3 <h2 style=";text-align:left;direction:ltr"> (F<h2 style=";text-align:left;direction:ltr"> input-low <h2 style=";text-align:left;direction:ltr"> )# ω=SE(H i +L i )# F output =Concat(ω*H i +(1-ω)*L i +H i +L i )# Among them, [] means grouping the feature maps by channels, F high Represents the high-dimensional feature map after grouping, F input-low Represents the initial low-dimensional feature map, Conv 3*3 represents a normal convolution layer with a convolution kernel of 3, F low Represents the low-dimensional feature map after grouping, H i and L i They represent a group of high-dimensional feature maps and low-dimensional feature maps after grouping, SE represents the channel attention module, ω represents the weight generated by the channel attention module, and Concat represents the concatenation of multiple feature maps by channel; F output Represents the output feature map of the FCSM module.

9. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 4, characterized in that: The NWD-CIoU regression loss function is implemented by parameter α nwd , perform weighted fusion of NWD and CIoU, and the calculation formula is as follows: Among them, cx a 、cy a 、w a 、h a are the center horizontal coordinate, center horizontal coordinate, width and height of the real frame, cx b 、cy b 、w b 、h b are the center horizontal coordinate, center horizontal coordinate, width and height of the prediction box respectively, and is the real box and the predicted box, W2 2 is the second-order Wasserstein distance between the real box and the predicted box, C is a constant, is the normalized second-order Wasserstein distance, is the NWD loss in the regression loss, α NWD It is a parameter to adjust CIoU loss and NWD loss, with a value of 0.

5. is the NWD-CIoU regression loss.

10. The method for detecting gynecological tumor lymph node metastasis based on edge feature extraction and fusion according to claim 1, characterized in that: In step S4, multiple indicators such as accuracy, recall, F1 value, average precision and model parameter quantity are used for evaluation and analysis, and a comprehensive analysis is performed using model inference results and heat maps.

Images