Medical image detection system and method

Abstract

The present invention relates to a medical image detection system and method, belonging to the fields of computer vision and image processing technology. An attention mechanism is added to the original conv4_3, fc7, conv6_2, and conv7_2 feature layers of an SSD neural network. Adjacent layers are then paired and feature fused using an early fusion approach to obtain an improved neural network model. The improved neural network model is then trained using medical image data, and finally, medical image detection is performed. The present invention adds an attention mechanism to the original SSD neural network and fuses the features after the attention mechanism. While the size of the feature layer remains unchanged after the attention mechanism, the features contained in the feature map outputted by the method focus more on the features of small objects, thereby enriching contextual features. This addresses the difficulty in obtaining richer shallow features and the problem of missed detection in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and image processing technology, and in particular to a medical image detection system and method combining an attention mechanism and feature fusion. Background Art

[0002] The type, morphology, and blood content of circulating tumor cells provide reliable information for tumor detection and are crucial for early cancer diagnosis, treatment monitoring, and prognosis assessment. While the number of cancer cells in the blood is typically very small, with white blood cells and red blood cells predominating, cancer cells differ morphologically from other cells. The diameters of the five types of red blood cells range from 6 to 20 microns, while circulating tumor cells are relatively large, ranging from 12 to 25 microns. Accurately distinguishing circulating tumor cells from normal cells requires consideration of the large number and small size of normal cells, requiring a corresponding improvement in the accuracy of small-target detection.

[0003] Among the current mainstream detection algorithms, the one-stage SSD algorithm shows good performance in both accuracy and speed. However, its disadvantage is that the detection accuracy of small targets is low. The reason is that the shallow features used to find small targets in the neural network structure are not nonlinear enough. In particular, some semantic information will be lost after the convolution layer and pooling layer, making it difficult to obtain richer shallow features. Secondly, there is the common problem of missed detection in target detection networks, especially when there are multiple overlapping small target positions. Even when there are a large number of preset anchor points and prior boxes, this problem still needs to be solved.

[0004] Therefore, this field urgently needs a technical solution that can obtain richer shallow features and solve the problem of missed detection. Summary of the Invention

[0005] The purpose of the present invention is to provide a medical image detection system and method, which adds an attention mechanism to the original SSD neural network and fuses the features after the attention mechanism, enriching the context features. The features contained in the feature map output by it pay more attention to the features of small targets, thereby solving the problem of difficulty in obtaining richer shallow features and missed detection in the existing technology.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A medical image detection system, comprising:

[0008] The attention mechanism addition module is used to add attention mechanisms to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network, respectively, to obtain four attention feature layers;

[0009] A feature layer combination module is used to combine the four attention feature layers in pairs between adjacent layers and perform feature fusion in an early fusion manner to obtain an improved neural network model;

[0010] A training module, configured to train the improved neural network model using medical image data to obtain a trained neural network model;

[0011] A detection module is used to perform medical image detection using the trained neural network model.

[0012] In some embodiments, the attention mechanism is a CBAM attention mechanism.

[0013] In some embodiments, the attention mechanism includes: a channel attention mechanism and a spatial attention mechanism.

[0014] In some embodiments, the attention mechanism is added to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers, specifically including:

[0015] Using the channel attention mechanism, the input features are subjected to global maximum pooling and global average pooling respectively to obtain maximum pooling results and average pooling results;

[0016] The maximum pooling result and the average pooling result are added through the shared fully connected layer in the SSD neural network, and the weight of each channel is obtained through the sigmoid activation function;

[0017] Multiplying the weight of each channel by the input feature to obtain a multiplied input feature;

[0018] Utilizing the spatial attention mechanism, taking the maximum and average values of the features for each feature point of the multiplied input features;

[0019] The maximum value and the average value of the feature of each feature point are stacked and passed through a convolution layer and a sigmoid activation function to obtain the weight of each feature point;

[0020] The weight of each feature point is multiplied by the input feature to obtain an attention feature layer.

[0021] In some embodiments, the stacking of the maximum value and the average value of the features of each feature point and the subsequent passing through a convolutional layer and a sigmoid activation function to obtain the weight of each feature point specifically includes:

[0022] The maximum value and the average value of the feature of each feature point are stacked and passed through a convolution layer with a step size of 1 and a sigmoid activation function to obtain the weight of each feature point.

[0023] In some embodiments, the four attention feature layers are combined in pairs between adjacent layers, and feature fusion is performed in an early fusion manner to obtain an improved neural network model, specifically including:

[0024] The conv4_3 feature layer after adding the attention mechanism is combined with the fc7 feature layer, the fc7 feature layer and the conv6_2 feature layer, and the conv6_2 feature layer and the conv7_2 feature layer to obtain three combined feature layers;

[0025] The number of channels of each combined feature layer is superimposed using the concat early fusion method to obtain three superimposed feature layers;

[0026] Each of the superimposed feature layers is passed through a convolution layer with a convolution kernel of 1 to obtain an improved neural network model.

[0027] In some embodiments, the combining of the conv4_3 feature layer and the fc7 feature layer after adding the attention mechanism specifically includes:

[0028] The fc7 feature layer after adding the attention mechanism is subjected to a deconvolution operation with a step size of 2 and a convolution kernel size of 1, and the fc7 feature layer after the deconvolution operation is combined with the conv4_3 feature layer after adding the attention mechanism.

[0029] In some embodiments, the step of training the improved neural network model using medical image data to obtain a trained neural network model specifically includes:

[0030] Treat white blood cells as normal cells;

[0031] mixing and culturing circulating tumor cells and the leukocytes;

[0032] acquiring a plurality of images containing the circulating tumor cells and the leukocytes;

[0033] Dividing the plurality of images containing the circulating tumor cells and the leukocytes into a training validation set and a test set;

[0034] The improved neural network model is trained using the training validation set and the test set to obtain a trained neural network model.

[0035] In some embodiments, after performing medical image detection using the trained neural network model, the method further includes:

[0036] Predict and count tumor cells and normal cells.

[0037] The present invention also provides a medical image detection method, the method comprising:

[0038] The attention mechanism is added to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers;

[0039] The four attention feature layers are combined in pairs between adjacent layers, and feature fusion is performed in an early fusion manner to obtain an improved neural network model;

[0040] Using medical image data to train the improved neural network model to obtain a trained neural network model;

[0041] The trained neural network model is used to perform medical image detection.

[0042] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0043] The medical image detection system and method provided by the present invention first adds an attention mechanism to the original conv4_3, fc7, conv6_2, and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers; then, the four attention feature layers are combined in pairs between adjacent layers and feature fusion is performed in an early fusion manner to obtain an improved neural network model; the improved neural network model is then trained using medical image data to obtain a trained neural network model; and finally, the trained neural network model is used to perform medical image detection. The present invention adds an attention mechanism to the original neural network and fuses the features after the attention mechanism. After the operation of the attention mechanism, the size of the feature layer does not change, but the features contained in the feature map outputted by it focus more on the features of small targets, thereby enriching the contextual features, thereby solving the problems of the difficulty in obtaining richer shallow features and missed detection in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a block diagram of a medical image detection system provided in Example 1 of the present invention.

[0046] Figure 2 This is a flowchart of the medical image detection method provided in Example 2 of the present invention.

[0047] Figure 3 Schematic diagram of adding an attention mechanism to an SSD neural network, provided in Example 2 of the present invention.

[0048] Figure 4 This is a schematic diagram comparing the average classification recognition accuracy before and after the improvement provided in the second embodiment of the present invention.

[0049] Figure 5 This is a schematic diagram of the distribution of interactive interface modules provided in the second embodiment of the present invention.

[0050] Figure 6 This is a schematic diagram of the interactive interface and detection results provided in Example 2 of the present invention.

[0051] Figure 7 This is a schematic diagram comparing the cell image detection results provided in Example 2 of the present invention.

[0052] Figure 8 This is a schematic diagram of the complete steps of the medical image detection method provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0053] 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] The purpose of the present invention is to provide a medical image detection system and method, which adds an attention mechanism to the original neural network and fuses the features after the attention mechanism, enriching the context features. The features contained in the feature map output by it pay more attention to the features of small targets, thereby solving the problem of difficulty in obtaining richer shallow features and missed detection in the existing technology.

[0055] Traditional detection methods are primarily based on template matching and machine learning, utilizing techniques such as sliding windows and support vector machines to locate specific regions. However, with the continuous development of deep learning and increased computing power, neural network architectures are able to effectively learn the rich features of different objects in an image, enabling better discrimination and exhibiting improved robustness and generalization. Deep learning detection algorithms are categorized into one-stage and two-stage approaches, depending on the number of neural networks involved. First-stage detection algorithms are faster and more efficient, while second-stage algorithms are computationally more complex but offer higher accuracy. Among the current mainstream detection algorithms, the one-stage SSD algorithm demonstrates good performance in both accuracy and speed. However, its drawbacks primarily include low accuracy for small object detection. This is due to the lack of nonlinearity in the shallow features used to locate small objects within the neural network architecture. In particular, after convolutional and pooling layers, some semantic information is lost, making it difficult to obtain rich shallow features. Furthermore, object detection networks suffer from the common problem of missed detection, particularly when multiple small objects overlap. This problem remains unresolved even with a large number of pre-set anchor points and prior bounding boxes. Therefore, it is necessary to improve the original SSD network structure. Therefore, the present invention introduces the technical means of attention mechanism and feature fusion to solve the above problems.

[0056] The present invention adds an attention mechanism to the original SSD network, and fuses the features after the attention mechanism, enriching the contextual features, and obtaining an improved SSD model that combines the attention mechanism and feature fusion. The improved SSD model adds the CBAM attention mechanism and feature fusion to the original SSD network structure, so that the network can adaptively focus on the important texture features of circulating tumor cells and normal cells and fully learn the contextual information between feature layers, thereby enhancing the model's detection accuracy for small targets and improving the model's generalization and robustness. The average recognition accuracy of the present invention on the self-made circulating tumor cell dataset has been increased from 85.00% to 91.54% compared to the original SSD model, with stronger generalization ability and higher small target detection ability. It also adds a cell counting function and designs an interactive interface, which can intuitively and conveniently select images and quickly generate detection results.

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Example 1:

[0059] like Figure 1 As shown, this embodiment provides a medical image detection system, the system comprising:

[0060] The attention mechanism adds a module M1, which is used to add attention mechanisms to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers.

[0061] In this embodiment, the attention mechanism is the CBAM attention mechanism, which includes: a channel attention mechanism and a spatial attention mechanism.

[0062] The attention mechanism is added to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers, including:

[0063] The channel attention mechanism is used to perform global maximum pooling and global average pooling on the input features to obtain maximum pooling results and average pooling results.

[0064] The maximum pooling result and the average pooling result are added through the shared fully connected layer in the SSD neural network, and the weight of each channel is obtained through the sigmoid activation function.

[0065] The weight of each channel is multiplied by the input feature to obtain a multiplied input feature.

[0066] The spatial attention mechanism is used to take the maximum value and average value of each feature point of the multiplied input feature.

[0067] The maximum value and the average value of the feature of each feature point are stacked and passed through a convolution layer with a step size of 1 and a sigmoid activation function to obtain the weight of each feature point.

[0068] The weight of each feature point is multiplied by the input feature to obtain an attention feature layer.

[0069] The feature layer combination module M2 is used to combine the four attention feature layers into pairs of adjacent layers and perform feature fusion in an early fusion manner to obtain an improved neural network model. The specific steps include:

[0070] The conv4_3 feature layer after adding the attention mechanism is combined with the fc7 feature layer, the fc7 feature layer and the conv6_2 feature layer, and the conv6_2 feature layer and the conv7_2 feature layer to obtain three combined feature layers.

[0071] The number of channels of each combined feature layer is superimposed respectively using the early fusion method of concat to obtain three superimposed feature layers.

[0072] Each of the superimposed feature layers is passed through a convolution layer with a convolution kernel of 1 to obtain an improved neural network model.

[0073] Among them, the conv4_3 feature layer after adding the attention mechanism is combined with the fc7 feature layer, specifically including:

[0074] The fc7 feature layer after adding the attention mechanism is subjected to a deconvolution operation with a step size of 2 and a convolution kernel size of 1, and the fc7 feature layer after the deconvolution operation is combined with the conv4_3 feature layer after adding the attention mechanism.

[0075] In this embodiment, the conv4_3, fc7, conv6_2 and conv7_2 feature layers after CBAM attention are combined in pairs, conv4_3 with fc7, c7 with conv6_2, and conv6_2 with conv7_2. The size of the feature map of the latter layer is expanded to the size of the previous layer through deconvolution, and then the number of channels of the front and back layers are superimposed through the early fusion method of concat. Finally, the convolution layer with a convolution kernel of 1*1 is used to reduce the number of channels to a size that can be passed to the SSD detection module, thereby realizing the feature fusion of the front and back layers, enriching the contextual semantic information, and obtaining an improved SSD model combining attention mechanism and feature fusion, that is, an improved neural network model.

[0076] The training module M3 is used to train the improved neural network model using medical image data to obtain a trained neural network model.

[0077] In this example, white blood cells are used as normal cells; circulating tumor cells and the white blood cells are mixed and cultured, and then several images containing the circulating tumor cells and the white blood cells are obtained. The images containing the circulating tumor cells and the white blood cells are divided into a training validation set and a test set. Finally, the improved neural network model is trained using the training validation set and the test set to obtain a trained neural network model.

[0078] As a specific embodiment, H1299 lung cancer cells are used to simulate circulating tumor cells in the medical image data, and white blood cells in the processed blood are used as normal cells; the data set has a total of 342 cell images, 90% of which are used as training and validation sets, and 10% are used as test sets. In the training and validation sets, 90% are used as training sets and 10% are used as validation sets to train the improved neural network model and the SSD neural network, respectively obtaining the training neural network model and the SSD neural network model.

[0079] The detection module M4 is used to perform medical image detection using the trained neural network model, and then predict and count tumor cells and normal cells.

[0080] In this embodiment, the training neural network model and the SSD neural network model are loaded, and the average accuracy of each category is calculated. Compared with the SSD neural network model, the training neural network model combining the attention mechanism and feature fusion achieves 91.54%, an improvement of 6.54%. This embodiment also designs a detection interface with interactive functions, which has the image module to be detected and a prediction module for loading the model generation results; the image module to be detected is used to select the microscope image to be detected from the file; the prediction module for loading the model generation results is used to select a model from the two trained model files and load the image passed in the previous step to obtain the prediction result, and identify the location information of circulating tumor cells and white blood cells and the number of cells within the field of view.

[0081] This embodiment discloses a medical image detection system that integrates attention mechanism and feature fusion. The CBAM attention mechanism is added to the original four feature layers of the SSD neural network: conv4_3, fc7, conv6_2, and conv7_2. The four feature layers after the attention mechanism are then paired together, and feature fusion is performed by increasing the number of channels for early fusion, resulting in an improved SSD model that combines attention mechanism and feature fusion. Both models are trained using a self-made cell dataset. Using the trained models, the average accuracy is calculated, and a detection interface is designed to predict and count circulating tumor cells and normal cells. The innovation of this embodiment lies in adding an attention mechanism to the original SSD network and fusing the features after the attention mechanism, enriching the contextual features. This results in an improved neural network model that combines attention mechanism and feature fusion. Compared to the original SSD model, the average recognition accuracy of this embodiment on the self-made circulating tumor cell dataset increased from 85.00% to 91.54%, demonstrating stronger generalization and improved small object detection capabilities. The addition of a cell counting function and an interactive interface enable intuitive and convenient image selection and rapid generation of detection results.

[0082] Example 2:

[0083] like Figure 2 As shown, this embodiment provides a medical image detection method, the method comprising:

[0084] S1. Add attention mechanisms to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers. Figure 3 As shown, the attention mechanism is the attention mechanism of CBAM.

[0085] The implementation of the CBAM attention mechanism is a combination of the channel attention mechanism and the spatial attention mechanism. Figure 3 As shown in (a); first, the channel attention mechanism performs global maximum pooling and global average pooling on the input features, uses a shared fully connected layer to add the two results, and then uses the sigmoid activation function to obtain the weight of each channel (each feature layer is composed of many channels, and the concept of the number of channels can be intuitively understood as Figure 3 (a) The length of the cuboid (the narrow one), and the number of feature points is the result of width*height. A feature layer is composed of length*width*height, and the weight after sigmoid can also be regarded as length*1*1. As long as the length is the same, the weight can be projected on each channel of the original feature layer accordingly). Then, the tensor containing the weights is kept the same as the number of channels of the input features, so that the weights can be multiplied by each feature point; secondly, the spatial attention mechanism takes the maximum value and average value of each feature point in the multiplication result, stacks the two results, and passes through a convolutional layer with a stride of 1 and a sigmoid activation function. , and then get the weight of each feature point. Similarly, under the premise of matching the number of channels, the tensor containing the weight is multiplied by the input feature, and finally a new feature layer is output; after the operation of these two attention mechanisms, the size of the feature layer does not change. The sizes of the four feature layers of conv4_3, fc7, conv6_2 and conv7_2 are still 38*38*512, 19*19*1024, 10*10*512 and 5*5*256 respectively, but the features contained in the feature maps they output are more focused on the features of small targets. The spatial attention mechanism in this embodiment takes the maximum and average values of the results obtained by the channel attention mechanism. If this spatial attention mechanism is used alone, as long as it is a feature layer, and the tensor and feature layer dimensions are kept matching in "length", it is actually still a multiplication of weights and features.

[0086] S2. The four attention feature layers are combined in pairs between adjacent layers, and feature fusion is performed in an early fusion manner to obtain an improved neural network model.

[0087] Feature fusion structure such as Figure 3As shown in (b), the four feature layers of conv4_3, fc7, conv6_2 and conv7_2 after CBAM attention are combined in pairs, conv4_3 with fc7, fc7 with conv6_2, and conv6_2 with conv7_2. First, a deconvolution operation with a stride of 2 and a kernel size of 1 is performed on fc7, so that its feature layer size matches that of conv4_3, with a channel number of 256. Then, using the early fusion method of concat, the number of channels is superimposed while ensuring that the size of conv4_3 and fc7 remains unchanged, resulting in a feature layer of 38*38*768. Finally, a convolution layer with a kernel of 1 is used to adjust the number of channels to 512, so that it can match the size of the first layer of the detection module in the SSD structure. Similarly, fc7 and conv6_2, and conv6_2 and conv7_2 are also fused in this way, and finally the size of 19*19*1024 and 10*10*512 are passed to the detection module respectively. In this embodiment, the number of channels is "long", but it is usually in the third dimension in the code (such as 768).

[0088] To verify the effectiveness of this structure, this example compares the average accuracy of the model without the attention mechanism using different feature fusion combinations. The test results are shown in Table 1.

[0089] Table 1 Experimental results using different feature fusion methods

[0090]

[0091] From the results in Table 1, we can see that the structure of directly fusing the first few feature layers and then passing them to the subsequent feature layers greatly affects the model's back propagation process, making it easy for the model to confuse the weights of the first few layers, causing the model's performance to deteriorate instead of improve. After adopting the strategy of directly passing the fused features into the detection module, through comparison, it is found that only when the features of two adjacent layers are fused, the model can use the contextual information between the feature layers to enrich the model's features, resulting in a slight improvement in its performance.

[0092] S3. Using medical image data to train the improved neural network model to obtain a trained neural network model.

[0093] In this embodiment, the self-made cell dataset is used to train the improved neural network model and the SSD neural network model respectively.

[0094] The homemade cell dataset uses H1299 lung cancer cells to simulate circulating tumor cells, and white blood cells from blood that has been treated with an ice-water bath and centrifuged to remove red blood cells are used as normal cells. Circulating tumor cells and white blood cells are mixed and cultured. The mixed cells are photographed using a confocal microscope, and a total of 57 images containing both types of cells are obtained. After random rotation, mirroring, and data augmentation, the dataset contains a total of 342 cell images; 90% of these images are used as a training validation set, and 10% as a test set. Within the training validation set, 90% is used as a training set and 10% as a validation set to train the SSD neural network, an improved SSD model combining an attention mechanism, and an improved SSD model combining an attention mechanism and feature fusion (i.e., an improved neural network model). The model is deployed on the GPU of the server for training and subsequent testing.

[0095] S4. Perform medical image detection using the trained neural network model.

[0096] In this embodiment, the trained model is used to calculate the mAP, and a detection interactive interface is designed to achieve the prediction and counting of circulating tumor cells and normal cells.

[0097] See also Figure 4 , the model was loaded on the test set for detection and the average accuracy of the two categories of circulating tumor cells and white blood cells was calculated. The SSD model can reach 85.00%, while the improved SSD model combining the attention mechanism and feature fusion reaches 91.54%. It can be seen intuitively that the improved SSD model has achieved different degrees of improvement for both categories.

[0098] See also Figure 5 The interactive detection interface is divided into a module for selecting images to be detected and a prediction module for loading models to generate results. In the module for selecting images to be detected, a blood microscope image containing circulating tumor cells is selected from a folder, and the results are generated after the model is selected. In the prediction module, a model to be loaded is selected. There are two detection models that can be loaded, the SSD model and the improved SSD model. After the model is selected, the prediction results are automatically generated and displayed on the interactive interface. The detection interface is as follows Figure 6 As shown, the model also adds a cell counting function, showing the number of tumor cells and white blood cells in the upper left corner of the image, allowing oncologists to intuitively and conveniently select images and quickly use the model for efficient detection during clinical diagnosis, accurately obtaining information on the number of tumor cells in patients.

[0099] See also Figure 7 , Figure 7The middle left and middle right are the original cell image, the detection results of the SSD model, and the detection results of the improved SSD model. It can be seen intuitively that the missed detection and false detection problems have been solved, and the confidence levels have also been improved, thus verifying that the effect of the improvement is more obvious.

[0100] The medical image detection method provided in this embodiment includes: 1. adding the CBAM attention mechanism to the original four feature layers of the SSD neural network, namely conv4_3, fc7, conv6_2 and conv7_2; 2. combining the four feature layers after adding attention in pairs, and performing feature fusion in a manner that increases the number of early fusion channels to obtain an improved SSD model that combines the attention mechanism and feature fusion; 3. using a self-made cell data set to train the two models separately; 4. using the trained model to calculate the average accuracy, and design a detection interaction interface to achieve the prediction and counting of circulating tumor cells and normal cells. The improved SSD model adds the CBAM attention mechanism and feature fusion to the original SSD network structure, so that the network can adaptively focus on the important texture features of circulating tumor cells and normal cells and fully learn the contextual information between feature layers, thereby enhancing the model's detection accuracy for small targets and improving the generalization and robustness of the model. The overall process step diagram is as follows. Figure 8 shown.

[0101] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0102] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A medical image detection system, characterized in that: The system comprises: The attention mechanism addition module is used to add attention mechanisms to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network, respectively, to obtain four attention feature layers; A feature layer combination module is used to combine the four attention feature layers in pairs between adjacent layers and perform feature fusion in an early fusion manner to obtain an improved neural network model; The four attention feature layers are combined in pairs between adjacent layers, and feature fusion is performed in an early fusion manner to obtain an improved neural network model, specifically including: The conv4_3 feature layer after adding the attention mechanism is combined with the fc7 feature layer, the fc7 feature layer and the conv6_2 feature layer, and the conv6_2 feature layer and the conv7_2 feature layer to obtain three combined feature layers; The number of channels of each combined feature layer is superimposed using the concat early fusion method to obtain three superimposed feature layers; Passing each of the superimposed feature layers through a convolution layer with a convolution kernel of 1 to obtain an improved neural network model; A training module, configured to train the improved neural network model using medical image data to obtain a trained neural network model; A detection module is used to perform medical image detection using the trained neural network model.

2. The medical image detection system according to claim 1, characterized in that: The attention mechanism is the CBAM attention mechanism.

3. The medical image detection system according to claim 1, wherein: The attention mechanism includes: channel attention mechanism and spatial attention mechanism.

4. The medical image detection system according to claim 3, characterized in that: The attention mechanism is added to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers, specifically including: Using the channel attention mechanism, the input features are subjected to global maximum pooling and global average pooling respectively to obtain maximum pooling results and average pooling results; The maximum pooling result and the average pooling result are added through the shared fully connected layer in the SSD neural network, and the weight of each channel is obtained through the sigmoid activation function; Multiplying the weight of each channel by the input feature to obtain a multiplied input feature; Utilizing the spatial attention mechanism, taking the maximum and average values of the features for each feature point of the multiplied input features; The maximum value and the average value of the feature of each feature point are stacked and passed through a convolution layer and a sigmoid activation function to obtain the weight of each feature point; The weight of each feature point is multiplied by the input feature to obtain an attention feature layer.

5. The medical image detection system according to claim 4, characterized in that: The weight of each feature point is obtained by stacking the maximum value and the average value of the feature of each feature point and passing them through a convolution layer and a sigmoid activation function, specifically including: The maximum value and the average value of the feature of each feature point are stacked and passed through a convolution layer with a step size of 1 and a sigmoid activation function to obtain the weight of each feature point.

6. The medical image detection system according to claim 1, characterized in that: The combination of the conv4_3 feature layer and the fc7 feature layer after adding the attention mechanism specifically includes: The fc7 feature layer after adding the attention mechanism is subjected to a deconvolution operation with a step size of 2 and a convolution kernel size of 1, and the fc7 feature layer after the deconvolution operation is combined with the conv4_3 feature layer after adding the attention mechanism.

7. The medical image detection system according to claim 1, characterized in that: The improved neural network model is trained using medical image data to obtain a trained neural network model, specifically comprising: Treat white blood cells as normal cells; mixing and culturing circulating tumor cells and the leukocytes; acquiring a plurality of images containing the circulating tumor cells and the leukocytes; Dividing the plurality of images containing the circulating tumor cells and the leukocytes into a training validation set and a test set; The improved neural network model is trained using the training validation set and the test set to obtain a trained neural network model.

8. The medical image detection system according to claim 1, wherein: After performing medical image detection using the trained neural network model, the method further includes: Predict and count tumor cells and normal cells.

9. A medical image detection method, characterized in that: The method comprises: The attention mechanism is added to the original conv4_3, fc7, conv6_2 and conv7_2 feature layers of the SSD neural network to obtain four attention feature layers; The four attention feature layers are combined in pairs between adjacent layers, and feature fusion is performed in an early fusion manner to obtain an improved neural network model; Using medical image data to train the improved neural network model to obtain a trained neural network model; Performing medical image detection using the trained neural network model; The four attention feature layers are combined in pairs between adjacent layers, and feature fusion is performed in an early fusion manner to obtain an improved neural network model, specifically including: The conv4_3 feature layer after adding the attention mechanism is combined with the fc7 feature layer, the fc7 feature layer and the conv6_2 feature layer, and the conv6_2 feature layer and the conv7_2 feature layer to obtain three combined feature layers; The number of channels of each combined feature layer is superimposed using the concat early fusion method to obtain three superimposed feature layers; Each of the superimposed feature layers is passed through a convolution layer with a convolution kernel of 1 to obtain an improved neural network model.

Images