Method, system, device and medium for detecting positive lymph nodes in medical images

Abstract

The application discloses a kind of detection method, system, equipment and medium of positive lymph node in medical image, it is related to medical image detection technical field, the method includes: collecting the preoperative CT image of sample individual;Preoperative CT image is input into image detection model, and the positive lymph node image features in different regions of preoperative CT image are extracted by the multi-scale feature extraction main network containing P1, P2, P3, P4, P5 layer;Multi-scale feature fusion module is used to fuse the multi-scale features extracted, and the region position of positive lymph node in medical image is identified according to the feature after fusion;The method shows superiority in automatically positioning, detecting different organ parts positive lymph node of medical image, improves multi-organ part, multi-target positive lymph node detection precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image detection, in particular to a method and system for detecting positive lymph nodes in a medical image, equipment and a medium. BACKGROUND

[0002] Imaging examination is the main method for diagnosing lymph node metastasis (LNM), and the determination of surgical range usually depends on imaging diagnosis. Therefore, accurately detecting lymph nodes and avoiding missing LNM are crucial for determining the lymph node radiotherapy target area.

[0003] Current lymph node detection mainly relies on clinical doctors to make diagnosis based on imaging features. For example, a scholar proposed an imaging-based radiomics signature strategy for tumor regions of preoperative MRI, which is used for binary classification prediction of bladder cancer lymph node metastasis (Wu S, et al, Development and Validation of an MRI-Based Radiomics Signature for the Preoperative Prediction of Lymph Node Metastasis in Bladder Cancer, EBioMedicine, 2018 Aug: 34: 76-84). The study extracts image features from the tumor region of preoperative MRI, and constructs a quantitative prediction model-nomogram through feature screening and logistic regression, which is used to determine whether the patient sample has lymph node metastasis.

[0004] However, the current image intelligent analysis-based method still stays at the level of binary classification discrimination of whether the tumor has lymph node metastasis, and has not gone deep into the positioning detection of tumor lymph node metastasis. The detection method used is difficult to further locate the positive lymph node position area in the preoperative image. SUMMARY

[0005] In view of the problem that the prior art cannot accurately locate the positive lymph node, the present application provides a method and system for detecting positive lymph nodes in a medical image, equipment and a medium, which constructs a detection framework for positive lymph nodes in different regional positions, attitudes and shapes, thereby solving the problems existing in the prior art.

[0006] A method for detecting positive lymph nodes in a medical image, comprising the following steps:

[0007] Collecting preoperative CT images of a sample individual;

[0008] The preoperative CT image is input into the image detection model, and the positive lymph node image features of different regions in the preoperative CT image are extracted by a multi-scale feature extraction backbone network including P1, P2, P3, P4 and P5 layers. Specifically, after the preoperative CT image is input into P1 and P2 for convolution operation, it is input into a bidirectional decomposition VSS module, respectively passes through different linear layers Linear and convolution layers, generates forward features and backward features, splices the forward features and the backward features to obtain output features, and after the output features of the P2 layer are operated in the P3 layer, they are input into a deformation convolution module DCNv3 to extract dominant features of positive lymph nodes in different regions. After the dominant features of positive lymph nodes in different regions extracted by the P3 layer are operated in the P4 layer and the P5 layer, the positive lymph node image features in different regions are obtained. A multi-scale feature fusion module is used to fuse the positive lymph node image features in different regions to obtain fused features.

[0009] The region containing the positive lymph node image is identified according to the fused features.

[0010] Further, the P1, P2, P3, P4 and P5 layers each include convolution operation, batch normalization operation and activation function; the activation function uses a LeakyHS activation function, which is represented as:

[0011]

[0012] Wherein, x represents the input feature.

[0013] Further, after the preoperative CT image is input into P1 and P2 for convolution operation, it is input into a bidirectional decomposition VSS module, respectively passes through different linear layers Linear and convolution layers, generates forward features and backward features, splices the forward features and the backward features to obtain output features, including the following steps:

[0014] Given a sequence of features x, after inputting it into the bidirectional decomposition VSS module, respectively passing through different linear layers Linear and convolution layers, forward features x f and backward features x b are generated, respectively represented as:

[0015]

[0016] Wherein, ForwardConv represents forward convolution operation, and BackwardConv represents inverse convolution operation.

[0017] The forward and backward state space models SSM are used to perform forward and backward scanning calculations on x f and x b respectively, and a same weight is used to obtain new forward features x'f and backward feature x' b ;

[0018]

[0019] The Gate of the forward and backward features is constructed based on the linear layer and the activation function LeakyHS respectively, and x' f and x' b are element-wise multiplied and then spliced to obtain the output feature x':

[0020] x' = Concat(x' f * LeakyHS(Linear(x)), x' b * LeakyHS(Linear(x)).

[0021] Further, the positive lymph node dominant features of different regions extracted by the P3 layer are operated at the P4 layer and the P5 layer to obtain the positive lymph node image features of different regions, which specifically include the following steps:

[0022] After the positive lymph node dominant features of different regions extracted by the P3 layer are operated at the P4 layer, they are input into the C2f module for splicing and compression;

[0023] After the spliced and compressed features are operated at the P5 layer, they are input into the C2f and SPFF modules to extract the positive lymph node image features of different regions.

[0024] Further, the multi-scale feature fusion module is used to fuse the extracted multi-scale features, which specifically includes the following steps:

[0025] The large-scale feature information of the P4 and P5 layers is integrated into the p3 layer by upsampling operation;

[0026] The feature information of different scales is further fused by the cross-stage partial bottleneck layer C2f with two convolution paths; meanwhile, the GAM module combining double channels and spatial attention is used to extract global attention to obtain the fused features.

[0027] It also includes using a loss function to distinguish the regions containing positive lymph node images; the loss function includes classification loss VFL, regression loss DFL and IoUL, which are respectively represented as:

[0028]

[0029] Wherein, alpha and gamma are hyperparameters, p γ is an adjustment factor, the value of q is the IOU between the prediction box and the real box, and p represents the classification probability;

[0030]

[0031] wherein, B gt represents a rectangular frame of the image position of the positive lymph node, B predict represents a predicted rectangular frame; Intersection represents the overlapping area between the real target frame and the predicted target frame, and Union represents the combined area of the two;

[0032] DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(y-y i )log(S i+1 ))

[0033] wherein, y i represents the value of the edge frame of the i-th prediction; y is the value of the real edge frame;

[0034] The total loss function Total Loss is represented as:

[0035]

[0036] wherein, n pos is the number of feature points allocated as positive samples, is the balance coefficient of the regression loss.

[0037] The application also includes a detection system for positive lymph nodes in medical images, comprising:

[0038] The acquisition module is configured to acquire preoperative CT images of a sample individual;

[0039] The feature extraction module is configured to input the preoperative CT images into an image detection model, extract image features of positive lymph nodes in different regions of the preoperative CT images through a multi-scale feature extraction backbone network comprising P1, P2, P3, P4 and P5 layers, and specifically configured to input the preoperative CT images into P1 and P2 for convolution operation, then input the preoperative CT images into a bidirectional decomposition VSS module, respectively pass through different linear layers Linear and convolution layers to generate forward features and backward features, splice the forward features and the backward features to obtain output features, input the output features of P2 into P3 for operation, then input the output features of P3 into a deformation convolution module DCNv3 to extract dominant features of positive lymph nodes in different regions, input the dominant features of positive lymph nodes in different regions extracted by P3 into P4 and P5 for operation to obtain image features of positive lymph nodes in different regions; and the multi-scale feature fusion module is configured to fuse the image features of positive lymph nodes in different regions to obtain fused features.

[0040] The recognition module is used to identify regions containing positive lymph node images based on the fused features.

[0041] The present invention also includes a computer device for detecting positive lymph nodes in medical images, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for detecting positive lymph nodes in medical images.

[0042] The present invention also includes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the method for detecting positive lymph nodes in medical images.

[0043] This invention provides a method, system, device, and medium for detecting positive lymph nodes in medical images, which has the following beneficial effects:

[0044] This invention designs a multi-scale feature extraction backbone network with a five-layer feature pyramid structure, and innovatively introduces VSS and DCNv3 modules in corresponding layers. The VSS module solves the problem of difficulty in extracting small target features of positive lymph nodes in the large background of CT images, while the DCNv3 module enhances the network's ability to identify positive lymph nodes in different body parts, regions, shapes, and postures. Combining these two modules effectively extracts semantic features, achieving the representation of positive lymph nodes in multiple regions, small sizes, and complex shapes and postures. Finally, a multi-scale feature fusion module is used to fuse the extracted multi-scale features to identify the regions containing positive lymph nodes in the image. This method, by constructing a detection framework for positive lymph nodes in different body parts, regions, and postures, demonstrates superiority in automatically locating, detecting, and visualizing positive lymph nodes in different parts of the patient, and can improve the prediction accuracy of multi-site and multi-regional lymph nodes. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method for detecting positive lymph nodes in medical images in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the detection method for positive lymph nodes in medical images in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the VSS module architecture in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the GAM module architecture in an embodiment of the present invention. Detailed Implementation

[0049] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0050] The present application proposes a multi-region positive lymph node detection technology. According to the biological characteristics of multi-region lymph nodes and their dominant signs in preoperative enhanced CT, a multi-scale deformable convolution feature extraction module is designed, which contains a multi-scale visual state space submodule and a deformable convolution network module, to fully explore the important features of the multiple, tiny and distorted targets (positive lymph nodes) hidden in the image region. On this basis, a feature / attention fusion module is designed, including multi-scale feature fusion and global attention fusion mechanism, to fully integrate feature information and network attention of different scales. Finally, a detection output module is designed to perform detection output at three resolution levels and select the optimal detection result. As shown in Figure 1 The detection method specifically includes the following steps:

[0051] S1, collecting the preoperative enhanced CT image of the patient sample.

[0052] S2, positive lymph node feature extraction; a multi-scale feature extraction backbone network is designed, and a total of five layers of feature pyramids P1, P2, P3, P4 and P5 are constructed, as shown in Figure 2 . P1 to P5 all contain convolution, batch normalization, activation function and other operation units. Among them, the LeakyHardSwish (Leaky HS) activation function is innovatively designed, as shown in formula (1), to effectively solve the problem of network gradient disappearance and enhance the convergence speed and model generalization performance. A bidirectional decomposition multi-scale visual state space (Visual State Space, VSS) submodule is designed after the P2 layer, with Gate operation, to fully solve the problem of feature extraction of small positive lymph node targets contained in the large background of CT images.

[0053]

[0054] The features output from the P1 operation unit enter the operation unit P2 of the next scale, which also contains convolution, batch normalization, and the activation function LeakyHS, to further extract high-dimensional information of the image. A VSS module is designed after the P2 operation unit, as shown in Figure 3 . Given a sequence of features x, when it is input into the bidirectional decomposition VSS module, it generates forward features x f and backward features x bAs shown in equation (2). Then, the above x f and x b are calculated by forward and backward scanning respectively, and share weights to obtain new forward feature x' f and backward feature x' b , as shown in equation (3). Finally, the gates of forward and backward features are constructed based on linear layer and activation function LeakyHS respectively, and are element-wise multiplied with x' f and x' b , and then spliced to form output feature x', as shown in equation (4). The purpose of using Gate operation is to weaken the direction bias introduced in forward / backward scanning and enhance the connection characteristics of different regions of the image.

[0055]

[0056] x' = Concat(x' f * LeakyHS(Linear(x)), x' b * LeakyHS(Linear(x)) (4)

[0057] P3, P4, and P5 layers are designed for higher scale feature extraction, and the layer structure is similar to that of P1 and P2 layers. After P3 layer, a Deformable Convolutional Networks v3 (DCNv3) module is innovatively introduced to fully exploit the important features of multiple, small, and twisted and deformed positive lymph nodes hidden in the image regions, as shown in equation (5), aiming to enhance the network's ability to recognize positive lymph nodes in different body parts, regional positions, shapes, and postures. Figure 2 DCNv3 is a variable convolution module that draws on the idea of separable convolution and separates the original convolution weight w g into a depth part and a point-by-point part. The depth part is responsible for the original position perception modulation scalar m gk , and the point-by-point part is the projection weight w shared between sampling points, aiming to enhance the network's ability to recognize positive lymph nodes in different body parts, regional positions, shapes, and postures, as shown in equation (5).

[0058]

[0059] where p0 is an arbitrary element of the input feature matrix; G represents the total number of aggregation groups, w g represents the position-independent projection weight of the group; represents the modulation scalar of the k-th sampling point in the g-th group, which is normalized along the dimension K by the softmax function; xg represents the input feature matrix of the g-th group; Δp gk is the input feature matrix of the g-th group at the grid sampling position p k corresponding offset. The C2f module is introduced after the P4 layer. The C2f module has two functions: one is feature aggregation, which effectively extracts and fuses multi-scale feature information by splicing the features output by different bottleneck modules and the original features input by the P4 layer, so as to realize effective representation of the detection target; the other is model compression, after the multi-scale feature fusion, the C2f adopts convolution operation, which can effectively compress the feature map, realize model lightweight, and effectively maintain the expression ability of the feature. The C2f and SPFF modules are introduced after the P5 layer, further realizing multi-scale feature extraction and model lightweight.

[0060] S3, multi-scale feature / global attention fusion; in order to fully integrate the different scale feature information output by the P3, P4 and P5 layers, first, the large scale feature information of the P4 and P5 layers is fused into the P3 layer by using the up-sampling operation, and then the features of different scales are further fused by the cross-stage partial bottleneck layer (Cross Stage Partial Bottom Neck with 2 Convolutions, C2f) with two convolution operations.

[0061] Since the captured low-level semantic features may not be sufficient to detect and distinguish the positive lymph node micro-target, the present application designs a multi-scale feature / global attention fusion module before detection, which effectively fuses the high-order semantic feature P5 layer and the low-order semantic features P3 and P4 by using up-sampling and C2f. On this basis, a GAM module with double channel and spatial attention integration is innovatively designed, as shown in Figure 4 which fully extracts global attention and better understands the relationship between the detection target and the global structure of the image, realizes multi-region and micro-target perception, maximally fuses channel and spatial attention, and realizes effective representation of the positive lymph node.

[0062] S4, multi-resolution detection output; the present application corresponds to the P3, P4 and P5 layers, and respectively designs corresponding detection outputs, so that the network can adapt to micro-targets of different scales and resolutions, improve the perception ability of multi-scale targets, and improve the detection accuracy of positive lymph nodes.

[0063] The loss function includes two branches: the classification loss Varifocal Loss (VFL), and the regression loss Distribution Focal Loss (DFL) and Intersection of Union Loss (IoUL), which correspond to the positive lymph node discrimination and accurate detection, respectively.

[0064]

[0065] wherein, both a and g are hyperparameters. a is a balance parameter, used to adjust the weight of positive and negative samples, while the adjustment factor p γ can reduce the influence of easy-to-classify samples on the loss, so that the model pays more attention to difficult-to-identify positive lymph node samples. For negative samples, when q = 0, p γ can be used to reduce the contribution of negative samples to the loss, while for positive samples (i.e. the case where q > 0), the value of q is the IOU between the predicted box and the real box. q is used to weight the positive samples, so that when the IOU of the positive sample is high, its contribution to the loss will also increase, so that the model pays more attention to high-quality positive samples, which is conducive to improving the detection accuracy of the model for positive lymph nodes.

[0066]

[0067] wherein, B gt and B predict represent the rectangular box where the positive lymph node is located, and the rectangular box predicted by the model, respectively.

[0068] DFL(S i ,S i+1 ) = - ((y i+1 -y) log (S i ) + (y-y i ) log (S i+1 )) (8)

[0069] Based on the above loss function, the total loss function Total Loss can be expressed as:

[0070]

[0071] wherein, n pos is the number of feature points assigned as positive samples, is the balance coefficient of the regression loss.

[0072] S5, positive lymph node detection result visualization; in order to improve the explainability of the model and increase the credibility of the model, Grad-CAM is introduced, which realizes the visualization process of model detection by using heat map method, at the same time evaluates the positioning ability of model detection, and at the same time brings the visual explanation of the model to the user.

[0073] The application constructs a positive lymph node feature extraction module, which introduces VSS and DCNv3 structures, can identify and remember the features of positive lymph nodes in different regions, and greatly filters out irrelevant information, thereby improving the mining ability of the network for important features of positive lymph nodes. After obtaining the important features of the positive lymph nodes, in order to further visualize the network prediction results and enhance their interpretability, a global attention mechanism method is introduced. This method can put global dimension interaction features in a reduced information dispersion, allowing the model to model the positive lymph nodes at different spatial scales, enabling the network to adapt to targets of different scales and resolutions, improving the perception of multi-scale targets, and ensuring the detection accuracy of the positive lymph nodes.

[0074] Based on the same inventive concept, the application further provides a system for detecting positive lymph nodes in medical images, comprising:

[0075] The acquisition module is configured to acquire preoperative CT images of a sample individual.

[0076] The feature extraction module is configured to input the preoperative CT images into an image detection model, extract image features of positive lymph nodes in different regions of the preoperative CT images through a multi-scale feature extraction backbone network comprising P1, P2, P3, P4 and P5 layers, and specifically configured to input the preoperative CT images into P1 and P2 for convolution operation, then input the preoperative CT images into a bidirectional decomposition VSS module, respectively pass through different linear layers Linear and convolution layers to generate forward features and backward features, splice the forward features and the backward features to obtain output features, input the output features of P2 into P3 for operation, then input the output features into a deformation convolution module DCNv3 to extract dominant features of positive lymph nodes in different regions, and input the dominant features of positive lymph nodes in different regions extracted by P3 into P4 and P5 for operation to obtain image features of positive lymph nodes in different regions.

[0077] The recognition module is configured to recognize regions containing positive lymph node images according to the fused features.

[0078] The application further provides a computer device for detecting positive lymph nodes in medical images, which comprises a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method for detecting positive lymph nodes in medical images.

[0079] The application further provides a readable storage medium, which stores a computer program, and the computer program comprises program instructions, and the program instructions are executed by a processor to perform the steps of the method for detecting positive lymph nodes in medical images.

[0080] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method of detecting positive lymph nodes in medical images, characterized in that, The method comprises the following steps: collecting preoperative CT images of sample individuals; The preoperative CT image is input into the image detection model, and the positive lymph node image features of different regions in the preoperative CT image are extracted through a multi-scale feature extraction backbone network including P 1、 P 2、 P 3、 P 4、 P 5 layers. Inputting preoperative CT images layer, P 2 layers are input into the bidirectional decomposition VSS module after convolution operation, respectively pass through different linear layers Linear and convolution layers, generate forward features and backward features, splice the forward features and the backward features to obtain output features; Specifically, the following steps are included: given a group of feature sequences , input them into the bidirectional decomposition VSS module, respectively pass through different linear layers Linear and convolution layers, generate forward features and backward features , respectively represented as: ; wherein, ForwardConv represents forward convolution operation, BackwardConv represents inverse convolution operation; the forward and backward state space models SSM are used to perform forward and backward scanning calculation on and respectively, and a same weight is adopted to obtain new forward features and backward features ; ; the gates of the forward and backward features are constructed based on the linear layers and the activation function LeakyHS respectively, and and are multiplied element by element to splice and obtain output features : ; Output features in P After computation in the 3-layer layer, the input to the deformable convolution module DCNv3 extracts the dominant features of positive lymph nodes in different regions. P The dominant features of positive lymph nodes in different regions extracted from 3 layers P 4 floors and P After five layers of computation, the image features of positive lymph nodes in different regions are obtained. Specifically, this includes the following steps: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] P After four layers of computation, the data is input into the C2f module for concatenation and compression; the concatenated and compressed features are then processed in... P After five layers of computation, the data are input into the C2f and SPFF modules to extract positive lymph node image features from different regions; wherein, the... P 1. P 2. P 3. P 4. P All five layers include convolution operations, batch normalization operations, and activation functions; a multi-scale feature fusion module is used to fuse the features of positive lymph node images from different regions to obtain the fused features; According to the fused features, the region containing the positive lymph node image is identified.

2. The method of claim 1, wherein the step of detecting the positive lymph node in the medical image is performed by using a deep learning model. The multi-scale feature fusion module is used to fuse the extracted multi-scale features, which specifically comprises the following steps: The scale feature information of the 5th layer is fused into the 4th layer by using an up-sampling operation. P 4、 P 5th layer is fused into the 4th layer by using an up-sampling operation. layer; Through the cross-stage partial bottleneck layer C2f with two convolution paths, the feature information of different scales is further fused; at the same time, the GAM module combining double channels and spatial attention is used to extract global attention, and the fused features are obtained.

3. The method of claim 1, wherein the step of detecting the positive lymph node in the medical image is performed by using a deep learning model. Also included is employing a loss function to discriminate the region containing the positive lymph node image; the loss function includes a classification loss VFL , a regression loss DFL and IoUL , respectively expressed as: wherein, and are hyperparameters, is a tuning factor, q the value of is the Intersection over Union between the predicted box and the ground truth box IOU , p is the target classification probability; wherein, a rectangular box representing the image position of the positive lymph node, a predicted rectangular box; Intersection an overlapping area between the real target box and the predicted target box, Union a combined area of the real target box and the predicted target box; wherein, represents the value of the bounding box for the first prediction; is the value of the bounding box for the real object. The total loss function is then Total Loss is represented as: wherein, is the number of feature points assigned as positive samples, is a balancing coefficient of the regression loss.

4. A system for detecting positive lymph nodes in medical images, characterized in that, It comprises: a collection module for collecting preoperative CT images of sample individuals; The feature extraction module is configured to input the preoperative CT image into an image detection model, and extract image features of positive lymph nodes in different regions in the preoperative CT image by using a multi-scale feature extraction backbone network including P 1、 P 2、 P 3、 P 4、 P 5 Inputting preoperative CT images layer, P After the convolution operation of the 2 layers, input the bidirectional decomposition VSS module, respectively pass through different linear layers Linear and convolution layers, generate forward features and backward features, splice the forward features and the backward features to obtain output features; Specifically, the following steps are included: Given a set of feature sequences , input the bidirectional decomposition VSS module, respectively pass through different linear layers Linear and convolution layers, generate forward features and backward features , respectively represented as: ; wherein, ForwardConv represents forward convolution operation, BackwardConv represents inverse convolution operation; using forward and backward state space models SSM to perform forward and backward scanning calculation on and respectively, and adopt the same weight to obtain new forward features and backward features ; ; respectively based on linear layers and activation function LeakyHS to construct gates of forward and backward features, and element-wise multiply and after splicing to obtain output features : ; The output features are input into the DCNv3 module to extract positive lymph node dominant features of different regions P 3 layers, and the positive lymph node dominant features of different regions extracted by the 3 layers are input into the DCNv3 module to extract positive lymph node dominant features of different regions P 3 layers, and the positive lymph node dominant features of different regions extracted by the 3 layers are input into the DCNv3 module to extract positive lymph node dominant features of different regions P 4 layers and P 5 layers, to obtain positive lymph node image features of different regions, which specifically include the following steps: inputting the positive lymph node dominant features of different regions into the DCNv3 module after operation of the 4 layers P 4 layers, and inputting the compressed features into the C2f module for splicing; inputting the spliced and compressed features into the DCNv3 module after operation of the 5 layers P 5 layers, to obtain positive lymph node image features of different regions; wherein the DCNv3 module includes 3 layers of convolutional operation, batch normalization operation and activation function P 1、 P 2、 P 3、 P 4、 P 5 layers all include convolutional operation, batch normalization operation and activation function; a multi-scale feature fusion module is used to fuse the positive lymph node image features of different regions to obtain fused features; An identification module is configured to identify a region containing a positive lymph node image according to the fused features.

5. A computer device for detecting positive lymph nodes in medical images, characterized in that, It comprises: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to realize the steps of the method for detecting positive lymph nodes in medical images according to any one of claims 1-3.

6. A readable storage medium characterized by, The readable storage medium stores a computer program, and the computer program comprises program instructions, which are executed by a processor to perform the steps of the method for detecting positive lymph nodes in medical images according to any one of claims 1-3.

Images