Intelligent Grading Method and System for Lumbar Disc Herniation Based on Global Attention CNN

Abstract

The present invention provides an intelligent grading method and system for lumbar disc herniation based on a global attention CNN. The method includes: Step 1: Locate the lumbar spine image to be measured to obtain an effective lumbar spine region, and then intercept the effective lumbar spine region image; Step 2: Improve the ResNeXt101 network to construct an intelligent grading network for lumbar disc herniation; Step 3: Train the intelligent grading network for lumbar disc herniation to obtain an intelligent grading network model for lumbar disc herniation; Step 4: Input the effective lumbar spine region image into the trained intelligent grading network model for lumbar disc herniation to obtain the grade information of lumbar disc herniation.

Description

Technical Field

[0001] The present invention relates to the technical fields of deep learning and computer vision, and particularly to an intelligent grading method and system for lumbar disc herniation based on global attention CNN. Background Art

[0002] Lumbar disc herniation is a chronic lumbar disease that is initially easily overlooked by people. The speed of early diagnosis is affected by many factors, including the referral time from the discovery of the condition to the hospital, the waiting time for an appointment with a specialist doctor, the time for an MRI scan, and the time for the analysis results to come out. The whole process may take several weeks to complete. In this way, the best treatment opportunity for the patient may be missed. Moreover, research shows that the probability of a patient returning to work and the required healthcare costs are directly proportional to the time of disability due to lumbar pain. Therefore, it is important to identify the cause of low back pain early to improve the patient's chance of recovery. At the same time, the accurate grading of the disease is crucial for the medical decision-making of physicians. However, traditional manual grading is affected by the experience level of physicians, resulting in differences in the grading results of different physicians, which in turn affects medical decision-making. Therefore, a fast and efficient lumbar disease grading method is needed.

[0003] At the same time, intelligent auxiliary diagnosis systems based on deep learning have been widely used in the medical field, and the diagnostic accuracy of many diseases has exceeded that of clinical physicians. The rapid development of the Internet, as well as disciplines such as computer vision, digital image processing, and artificial intelligence, and the integration of various new technologies with medicine have laid a solid technical foundation for the research and development of an automated diagnosis system for lumbar diseases. The research and development of an automated diagnosis system for lumbar diseases based on deep learning can not only reduce the increasing workload of radiologists, but also improve the utilization value of medical image data and reduce the misdiagnosis rate of clinical diagnosis, which is of great significance for clinical applications. Summary of the Invention[[ID=??]]

[0004] In order to overcome the problems that the grading of lumbar disc herniation is highly dependent on manual experience and is relatively labor-intensive, the present invention provides an intelligent grading method and system for lumbar disc herniation based on global attention CNN. While extracting high-level image semantic features, the present invention can also focus on learning low-level image detail information, and thus can accurately, efficiently, and quickly achieve the grading task of lumbar disc herniation.

[0005] On the one hand, the present invention provides an intelligent grading method for lumbar disc herniation based on global attention CNN, including:

[0006] Step 1: Locate the lumbar image to be measured to obtain an effective lumbar region, and then intercept the effective lumbar region image;

[0007] Step 2: Improve the ResNeXt101 network to construct an intelligent lumbar disc herniation grading network, specifically including: adding a spatial attention module attention1 at the input end of the ResNeXt101 network, adding a spatial attention module attention2 after layer4 in the ResNeXt101 network, and adding a main classifier after attention2; adding an auxiliary classifier 1 and using the feature map output by layer2 in the ResNeXt101 network as the input of the auxiliary classifier 1, adding an auxiliary classifier 2 and using the feature map output by layer3 in the ResNeXt101 network as the input of the auxiliary classifier 2; the sum of the outputs of the main classifier, the output of the auxiliary classifier 1, and the output of the auxiliary classifier 2 is used as the input of the Softmax classifier, and the Softmax classifier outputs the grade information of lumbar disc herniation;

[0008] Step 3: Train the intelligent lumbar disc herniation grading network to obtain an intelligent lumbar disc herniation grading network model;

[0009] Step 4: Input the effective lumbar region image into the trained intelligent lumbar disc herniation grading network model to obtain the grade information of lumbar disc herniation.

[0010] Further, the data processing process of the attention1 specifically includes:

[0011] Receiving the input feature map map1_1, evenly dividing the feature map map1_1 into 256 small blocks, and splicing them in the channel dimension to obtain the feature map map1_2. The feature map map1_2 is successively passed through global average pooling, a convolutional layer, and a Sigmoid activation function to obtain the feature map map1_3. The feature map map1_3 is multiplied element-wise with the feature map map1_2 to obtain the feature map map1_4. The result after successively passing the feature map map1_4 through a splicing operation, a convolutional layer, and a Sigmoid activation function is multiplied element-wise with the feature map map1_1 to obtain the feature map map1_6. The feature map map1_6 is added to the feature map map1_1 to obtain the feature map map1_7. Finally, the result of splicing the feature maps map1_1, map1_6, and map1_7 in the channel dimension is used as the output of attention1.

[0012] Further, the data processing process of the attention2 specifically includes:

[0013] Receive the feature map map2_1 output by layer4 in the ResNeXt101 network. After passing through the convolutional layer, the feature map map2_1 obtains the feature map map2_2. Then, the result after passing the feature map map2_2 through the Reshape operation, convolutional layer, Reshape operation, and Sigmoid activation function in sequence is multiplied element-wise with the feature map map2_1 to obtain the feature map map2_5. The result of adding the feature map map2_5 and the feature map map2_1 is used as the output of attention2.

[0014] Furthermore, the data processing process of the main classifier specifically includes:

[0015] Receive the feature map map2_6 output by attention2. After passing through global average pooling and the Reshape operation in sequence, the feature map map2_6 obtains the feature vector F2_1. After passing the feature vector F2_1 through a fully connected layer FC, the feature vector F2_2 is obtained. The feature vector F2_2 is used as the output of the main classifier.

[0016] Furthermore, the data processing process of the auxiliary classifier 1 specifically includes:

[0017] Receive the feature map map3_1 output by layer2 in the ResNeXt101 network. After passing through global average pooling and the Reshape operation in sequence, the feature map map3_1 obtains the feature vector F3_1. After passing the feature vector F3_1 through a fully connected layer FC, the feature vector F3_2 is obtained. The feature vector F3_2 is multiplied element-wise with a trainable weight vector W1 to obtain the feature vector F3_3. The feature vector F3_3 is used as the output of the auxiliary classifier 1.

[0018] Furthermore, the data processing process of the auxiliary classifier 2 specifically includes:

[0019] Receive the feature map map4_1 output by layer3 in the ResNeXt101 network. After passing through global average pooling and the Reshape operation in sequence, the feature map map4_1 obtains the feature vector F4_1. After passing the feature vector F4_1 through a fully connected layer FC, the feature vector F4_2 is obtained. The feature vector F4_2 is multiplied element-wise with a trainable weight vector W2 to obtain the feature vector F4_3. The feature vector F4_3 is used as the output of the auxiliary classifier 2.

[0020] Furthermore, in step 1, use the trained object detection model to localize the lumbar spine image to be measured.

[0021] Furthermore, step 1 also includes: performing image size normalization processing on the effective lumbar spine region image;

[0022] Correspondingly, step 4 is: inputting the effective lumbar region image after image size normalization processing into the trained intelligent grading network model for lumbar disc herniation to obtain the grade information of lumbar disc herniation.

[0023] Furthermore, the lumbar spine image to be measured is a lumbar spine MRI image in the T2 axial plane.

[0024] On the other hand, the present invention provides an intelligent grading system for lumbar disc herniation based on a global attention CNN, including:

[0025] An image preprocessing module, which is used to locate the lumbar spine image to be measured to obtain an effective lumbar region, and then intercept the effective lumbar region image;

[0026] An intelligent grading network module, which improves the ResNeXt101 network to construct an intelligent grading network for lumbar disc herniation. Specifically, it includes: adding a spatial attention module attention1 at the input end of the ResNeXt101 network, adding a spatial attention module attention2 after layer4 in the ResNeXt101 network, and adding a main classifier after attention2; adding an auxiliary classifier 1 and using the feature map output by layer2 in the ResNeXt101 network as the input of the auxiliary classifier 1, adding an auxiliary classifier 2 and using the feature map output by layer3 in the ResNeXt101 network as the input of the auxiliary classifier 2; the sum of the outputs of the main classifier, the output of the auxiliary classifier 1, and the output of the auxiliary classifier 2 is used as the input of the Softmax classifier, and the Softmax classifier outputs the grade information of lumbar disc herniation;

[0027] A training module, which is used to train the intelligent grading network for lumbar disc herniation to obtain an intelligent grading network model for lumbar disc herniation;

[0028] An intelligent grading module, which is used to input the effective lumbar region image corresponding to the lumbar spine image to be measured into the trained intelligent grading network model for lumbar disc herniation to obtain the grade information of lumbar disc herniation.

[0029] Advantages of the present invention:

[0030] 1. The present invention further optimizes on the basis of the ResNeXt101 network, and proposes a new network structure suitable for intelligent grading of lumbar disc herniation; and on this basis, a spatial attention module based on global information interaction is also designed, making the network learning more focused. In addition, two auxiliary classifiers are also designed. The present invention is applicable to the grading tasks of lumbar disc herniation in different positions, and has the characteristics of high accuracy and strong generalization ability.

[0031] 2. The spatial attention module attention1 based on global information proposed by the present invention adopts a strategy of dividing the image into blocks at the input end of the network. Calculating the importance of each block can reduce the number of parameters. When calculating the importance of each block, the information of all other blocks is combined, making the calculation method of spatial attention more reasonable.

[0032] 3. The spatial attention module attention2 based on global information interaction proposed by the present invention first passes through a convolutional layer with a convolution kernel of 1 at the input to integrate the multi-channel feature map into a single-channel feature map. The advantage of this is that while reducing the number of parameters, other channel information will not be lost. The feature map is unfolded in the channel dimension and then passes through a convolutional layer with a convolution kernel size of 1×1, realizing information interaction with other pixel points when calculating the importance of a single pixel point.

[0033] 4. Add two auxiliary classifiers on the basis of the ResNeXt101 network, and let the output of each auxiliary classifier be dot-multiplied with a 1D weight vector of length 11 with trainable parameters, and then add the outputs of the three classifiers, which can further improve the accuracy of the model. Description of the Drawings

[0034] Figure 1 It is one of the schematic flowcharts of the intelligent grading method for lumbar disc herniation based on global attention CNN provided by the embodiment of the present invention;

[0035] Figure 2 It is the second schematic flowchart of the intelligent grading method for lumbar disc herniation based on global attention CNN provided by the embodiment of the present invention;

[0036] Figure 3 It is the schematic diagram of image size normalization provided by the embodiment of the present invention;

[0037] Figure 4 It is the schematic diagram of the structure of the spatial attention module attention1 based on global information provided by the embodiment of the present invention;

[0038] Figure 5 It is the schematic diagram of the structure of the spatial attention module attention2 based on global information provided by the embodiment of the present invention;

[0039] Figure 6 It is the schematic diagram of the structure of the main classifier provided by the embodiment of the present invention;

[0040] Figure 7 It is the schematic diagram of the structure of the auxiliary classifier 1 provided by the embodiment of the present invention;

[0041] Figure 8 It is the schematic diagram of the structure of the auxiliary classifier 2 provided by the embodiment of the present invention. Specific Embodiments

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] In the present invention, both the lumbar spine images to be measured and the training images used when training the network model are lumbar spine MRI images in the T2 axial plane.

[0044] Embodiment 1

[0045] As Figure 1 shown, the embodiment of the present invention provides an intelligent grading method for lumbar disc herniation based on a global attention CNN, including the following steps:

[0046] S101: Locate the lumbar spine image to be measured to obtain an effective lumbar spine region, and then intercept the effective lumbar spine region image;

[0047] Specifically, in practical applications, since lumbar spine MRI images, in addition to the images of the lumbar spine region, also contain a large range of background information, this background information will introduce noise, thereby increasing the difficulty of intelligent grading of lumbar disc herniation. Therefore, it is necessary to locate and intercept the effective lumbar spine region image. Manual positioning can be used, or a trained object detection model can be used for automatic positioning. In this embodiment, in order to quickly and accurately locate the effective region, the R-CNN detection model is used as the object detection model to mine the effective region in the original lumbar spine image to be measured, which can reduce the workload of manual labor and improve work efficiency at the same time.

[0048] In addition, considering that the sizes of the images intercepted from the original lumbar spine image to be measured are different, while the input of the convolutional neural network requires a unified image size (as Figure 3 shown), so this step also includes: performing image size normalization processing on the intercepted effective lumbar spine region image. In this embodiment, when adjusting the image size, the bilinear interpolation algorithm is used to scale the length and width of the image proportionally, and the blank parts are filled with black (the pixel value is set to 0). In this way, the image after size normalization will not be distorted and can meet the input requirements of the neural network, effectively improving the accuracy of lumbar disc herniation grading.

[0049] S102: Improve the ResNeXt101 network to construct an intelligent grading network for lumbar disc herniation;

[0050] Specifically, as Figure 2 shown, the improvements to the ResNeXt101 network specifically include: adding a spatial attention module attention1 at the input end of the ResNeXt101 network, adding a spatial attention module attention2 after layer4 in the ResNeXt101 network, and adding a main classifier after attention2; adding an auxiliary classifier 1 and using the feature map output by layer2 in the ResNeXt101 network as the input of the auxiliary classifier 1, adding an auxiliary classifier 2 and using the feature map output by layer3 in the ResNeXt101 network as the input of the auxiliary classifier 2; the sum of the outputs of the main classifier, the auxiliary classifier 1, and the auxiliary classifier 2 is used as the input of the Softmax classifier, and the Softmax classifier outputs the grade information of lumbar disc herniation.

[0051] S103: Train the lumbar disc herniation intelligent grading network to obtain a lumbar disc herniation intelligent grading network model;

[0052] S104: Input the effective lumbar region image into the trained lumbar disc herniation intelligent grading network model to obtain the grade information of lumbar disc herniation.

[0053] In the embodiment of the present invention, by adding the spatial attention modules attention1 and attention2, while extracting high-level image semantic features, it is also possible to focus on learning low-level image detail information. Then, the sum of the results of adding the three classifiers, namely the main classifier, the auxiliary classifier 1, and the auxiliary classifier 2, is used as the input of the Softmax classifier. Finally, the Softmax classifier outputs the grade information of lumbar disc herniation, so that the grading task of lumbar disc herniation can be accurately, efficiently, and quickly achieved.

[0054] Embodiment 2

[0055] In the above Embodiment 1, attention1 and attention2 can adopt existing spatial attention modules. However, considering that the existing spatial attention module only depends on the information of a certain pixel point in different channels when calculating the importance of a pixel point, when the number of channels is small, the performance improvement of calculating spatial attention for the entire network model is limited. Moreover, the calculation method of the existing spatial attention module only involves the features between channels and does not involve the interaction of global information. However, the importance of a pixel point in an image not only depends on the performance of the pixel point in different channels, but more importantly, on the relationship between the pixel point and its surrounding and other pixel points. Therefore, the embodiment of the present invention proposes a spatial attention module based on global information interaction, making the calculation method of spatial attention more scientific.

[0056] As shown Figure 4 in the figure, the data processing process of the spatial attention module attention1 based on global information interaction provided by the embodiments of the present invention specifically includes:

[0057] Receiving the input feature map map1_1, evenly dividing the feature map map1_1 into 256 small blocks, and concatenating them in the channel dimension to obtain the feature map map1_2. The feature map map1_2 is sequentially passed through global average pooling, a convolutional layer, and a Sigmoid activation function to obtain the feature map map1_3. The feature map map1_3 is multiplied element-wise with the feature map map1_2 to obtain the feature map map1_4. The result after sequentially passing the feature map map1_4 through a concatenation operation, a convolutional layer, and a Sigmoid activation function is multiplied element-wise with the feature map map1_1 to obtain the feature map map1_6. The feature map map1_6 is added to the feature map map level to obtain the feature map map1_7. Finally, the result of concatenating the feature maps map1_1, map1_6, and map1_7 in the channel dimension is used as the output of attention1.

[0058] As an example, assume that the size of the image after size normalization is 96×160. Under this setting, the processing process of attention1 is as follows: First, the 96×160-sized image map1_1 is evenly divided into 256 small blocks of size 6×10, and concatenated in the channel dimension to obtain an image map1_2 of size 256×6×10. Then, map1_2 undergoes global average pooling to obtain a feature map map1_8 of size 256×1×1. Map1_8 passes through a convolutional layer with a convolutional kernel of 1×1 to obtain a feature map map1_9 of size 256×1×1. Map1_9 passes through a Sigmoid activation function to obtain a feature map map1_3 of size 256×1×1. Then, map1_2 is multiplied element-wise with map1_3 to obtain a feature map map1_4 of size 256×6×10. Map1_4 is concatenated to obtain a feature map map1_5 of size 1×96×160. Map1_5 passes through a convolutional layer with a convolutional kernel of 1×1 to obtain a feature map map1_10 of size 1×96×160. Map1_10 passes through a Sigmoid activation function and is multiplied element-wise with map1_1 to obtain a feature map map1_6 of size 1×96×160. Map1_6 is added to map1_1 to obtain a feature map map1_7 of size 1×96×160. Finally, map1_1, map1_6, and map1_7 are concatenated in the channel dimension to obtain a feature map map1_11 of size 3×96×160. Map1_11 is used as the output of attention1.

[0059] In this embodiment, the image input to attention1 is first divided into blocks, and all the blocks are concatenated in the channel dimension. Then, a convolutional layer with a convolution kernel of 1×1 is used to realize the information interaction between different blocks. Among them, the advantage of dividing the image into blocks is that it can reduce the number of parameters required for calculating the spatial importance.

[0060] As Figure 5 shown, the data processing process of the spatial attention module attention2 based on global information interaction provided by the embodiment of the present invention specifically includes:

[0061] Receive the feature map map2_1 output by layer4 in the ResNeXt101 network. After passing through a convolutional layer, the feature map map2_2 is obtained. Then, the result after the feature map map2_2 passes through the Reshape operation, convolutional layer, Reshape operation, and Sigmoid activation function in sequence is multiplied by the feature map map2_1 to obtain the feature map map2_5. The result after adding the feature map map2_5 and the feature map map2_1 is used as the output of attention2.

[0062] As an example, map1_11, as the output of attention1, is input to the ResNeXt101 network for processing. The layer4 in the ResNeXt101 network finally outputs a feature map map2_1 with a size of 2048×12×20; map2_1 is input to attention2. After passing through a convolutional layer with a convolution kernel of 1×1, a feature map map2_2 with a size of 1×12×20 is obtained. map2_2 passes through the Reshape operation to obtain a feature map map2_3 with a size of 240×1×1. map2_3 passes through a convolutional layer with a convolution kernel of 1×1 to obtain a feature map map2_7 with a size of 240×1×1. map2_7 passes through the Reshape operation to obtain a feature map map2_4 with a size of 1×12×20. map2_4 passes through the Sigmoid activation function and is multiplied by map2_1 to obtain a feature map map2_5 with a size of 2048×12×20. Finally, map2_5 and map2_1 are added to obtain a feature map map2_6 with a size of 2048×12×20. map2_6 is used as the output of attention2.

[0063] In this embodiment, the feature map input to attention2 first passes through a convolutional layer with a 1×1 convolutional kernel, which can integrate the multi-channel feature map into a single-channel feature map. The advantage of this is that while reducing the number of parameters, other channel information will not be lost. Then, the feature map is unfolded in the channel dimension and then passes through a convolutional layer with a convolutional kernel size of 1×1, so that information interaction at each point can be achieved.

[0064] Embodiment 3

[0065] Based on the above embodiments, in order to further improve the hierarchical accuracy of the entire network model, in the embodiment of the present invention, as Figure 6 shown, the data processing process of the main classifier specifically includes: receiving the feature map map2_6 output by attention2. The feature map map2_6 sequentially passes through global average pooling and Reshape operations to obtain a feature vector F2_1. The feature vector F2_1 passes through a fully connected layer FC to obtain a feature vector F2_2, and the feature vector F2_2 is used as the output of the main classifier.

[0066] As an example, the feature map map2_6 with a size of 2048×12×20 output by attention2 passes through global average pooling to obtain a feature map map2_8 with a size of 2048×1×1. Map2_8 passes through a Reshape operation to obtain a 1D feature vector F2_1 with a length of 2048. F2_1 passes through a fully connected layer to obtain a 1D feature vector F2_2 with a length of 11, and F2_2 is used as the output of the main classifier.

[0067] As Figure 7 shown, the data processing process of the auxiliary classifier 1 specifically includes: receiving the feature map map3_1 output by layer2 in the ResNeXt101 network. The feature map map3_1 sequentially passes through global average pooling and Reshape operations to obtain a feature vector F3_1. The feature vector F3_1 passes through a fully connected layer FC to obtain a feature vector F3_2. The feature vector F3_2 is multiplied by a parameter-trainable weight vector W1 to obtain a feature vector F3_3, and the feature vector F3_3 is used as the output of the auxiliary classifier 1.

[0068] As an example, map1_11, which is the output of attention1, is input into the ResNeXt101 network for processing. In the ResNeXt101 network, layer2 finally outputs a feature map map3_1 with a size of 512×48×80. Map3_1 undergoes global average pooling to obtain a feature map map3_2 with a size of 512×1×1. Map3_2 undergoes a Reshape operation to obtain a 1D feature vector F3_1 with a length of 512. F3_1 passes through a fully connected layer to obtain a 1D feature vector F3_2 with a length of 11. F3_2 is dot-multiplied with a 1D weight vector W1 with a length of 11 whose parameters are trainable to obtain a 1D feature vector F3_3 with a length of 11. F3_3 is used as the output of the auxiliary classifier 1.

[0069] As Figure 8 shown, the data processing process of the auxiliary classifier 2 specifically includes: receiving the feature map map4_1 output by layer3 in the ResNeXt101 network. The feature map map4_1 sequentially undergoes global average pooling and a Reshape operation to obtain a feature vector F4_1. The feature vector F4_1 passes through a fully connected layer FC to obtain a feature vector F4_2. The feature vector F4_2 is dot-multiplied with a weight vector W2 with trainable parameters to obtain a feature vector F4_3. The feature vector F4_3 is used as the output of the auxiliary classifier 2.

[0070] As an example, map1_11, which is the output of attention1, is input into the ResNeXt101 network for processing. In the ResNeXt101 network, layer3 finally outputs a feature map map4_1 with a size of 1024×24×40. Map4_1 undergoes global average pooling to obtain a feature map map4_2 with a size of 1024×1×1. Map4_2 undergoes a Reshape operation to obtain a 1D feature vector F4_1 with a length of 1024. F4_1 passes through a fully connected layer to obtain a 1D feature vector F4_2 with a length of 11. F4_2 is dot-multiplied with a 1D weight vector W2 with a length of 11 whose parameters are trainable to obtain a 1D feature vector F4_3 with a length of 11. F4_3 is used as the output of the auxiliary classifier 2.

[0071] In the embodiments of the present invention, both of the two auxiliary classifiers respectively have a 1D weight vector with a length of 11 whose parameters are trainable, and the output of each auxiliary classifier is dot-multiplied with a 1D weight vector with a length of 11 (the same as the number of categories) whose parameters are trainable, which can further improve the accuracy of the model.

[0072] Example 4

[0073] Corresponding to the above method, an embodiment of the present invention further provides a lumbar disc herniation intelligent grading system based on global attention CNN, including: an image preprocessing module, an intelligent grading network module, a training module, and an intelligent grading module;

[0074] Among them, the image preprocessing module is used to locate the lumbar spine image to be measured to obtain an effective lumbar spine area, and then intercept the effective lumbar spine area image. The intelligent grading network module is used to improve the ResNeXt101 network to construct a lumbar disc herniation intelligent grading network, specifically including: adding a spatial attention module attention1 at the input end of the ResNeXt101 network, adding a spatial attention module attention2 after layer4 in the ResNeXt101 network, and adding a main classifier after attention2; adding an auxiliary classifier 1 and using the feature map output by layer2 in the ResNeXt101 network as the input of the auxiliary classifier 1, adding an auxiliary classifier 2 and using the feature map output by layer3 in the ResNeXt101 network as the input of the auxiliary classifier 2; the sum of the outputs of the main classifier, the output of the auxiliary classifier 1, and the output of the auxiliary classifier 2 is used as the input of the Softmax classifier, and the Softmax classifier outputs the grade information of lumbar disc herniation. The training module is used to train the lumbar disc herniation intelligent grading network to obtain a lumbar disc herniation intelligent grading network model. The intelligent grading module is used to input the effective lumbar spine area image corresponding to the lumbar spine image to be measured into the trained lumbar disc herniation intelligent grading network model to obtain the grade information of lumbar disc herniation.

[0075] It should be noted that the system provided by the embodiment of the present invention is to implement the above method embodiment, and its functions can be specifically referred to the above method embodiment, which will not be elaborated here.

[0076] The intelligent grading method and system for lumbar disc herniation based on global attention CNN provided by the present invention further optimizes on the basis of the ResNeXt101 network, and proposes a new network structure suitable for intelligent grading of lumbar disc herniation; and, on this basis, a spatial attention module based on global information interaction is also designed to make the network learning more focused. In addition, two auxiliary classifiers are designed, and the output of each auxiliary classifier is dot-multiplied with a trainable weight vector with a 1D length equal to the number of categories, which can further improve the accuracy. Therefore, the CNN lumbar disc herniation grading method and system implemented by the present invention have the characteristics of high accuracy.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An intelligent grading method for lumbar disc herniation based on global attention CNN, characterized in that Including: Step 1: Locate the lumbar spine image to be measured to obtain an effective lumbar spine area, and then intercept the effective lumbar spine area image; Step 2: Improve the ResNeXt101 network to construct an intelligent grading network for lumbar disc herniation, specifically including: adding a spatial attention module attention1 at the input end of the ResNeXt101 network, adding a spatial attention module attention2 after layer4 in the ResNeXt101 network, and adding a main classifier after attention2; adding an auxiliary classifier 1 and using the feature map output by layer2 in the ResNeXt101 network as the input of the auxiliary classifier 1, adding an auxiliary classifier 2 and using the feature map output by layer3 in the ResNeXt101 network as the input of the auxiliary classifier 2; the sum of the outputs of the main classifier, the output of the auxiliary classifier 1, and the output of the auxiliary classifier 2 is used as the input of the Softmax classifier, and the Softmax classifier outputs the grade information of lumbar disc herniation; Step 3: Train the intelligent grading network for lumbar disc herniation to obtain an intelligent grading network model for lumbar disc herniation; Step 4: Input the effective lumbar spine area image into the trained intelligent grading network model for lumbar disc herniation to obtain the grade information of lumbar disc herniation.

2. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, wherein The data processing process of the attention1 specifically includes: Receiving the input feature map map1_1, evenly dividing the feature map map1_1 into 256 small blocks, and splicing them in the channel dimension to obtain the feature map map1_2. The feature map map1_2 is successively passed through global average pooling, a convolutional layer, and a Sigmoid activation function to obtain the feature map map1_3. The feature map map1_3 is multiplied by the feature map map1_2 point by point to obtain the feature map map1_4. The result after successively passing the feature map map1_4 through a splicing operation, a convolutional layer, and a Sigmoid activation function is multiplied by the feature map map1_1 to obtain the feature map map1_6. The feature map map1_6 is added to the feature map map1_1 to obtain the feature map map1_7. Finally, the result of splicing the feature maps map1_1, map1_6, and map1_7 in the channel dimension is used as the output of attention1.

3. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, wherein, The data processing process of the attention2 specifically includes: Receiving the feature map map2_1 output by layer4 in the ResNeXt101 network, the feature map map2_1 is passed through a convolutional layer to obtain the feature map map2_2. Then, the result after successively passing the feature map map2_2 through a Reshape operation, a convolutional layer, a Reshape operation, and a Sigmoid activation function is multiplied by the feature map map2_1 point by point to obtain the feature map map2_5. The result of adding the feature map map2_5 to the feature map map2_1 is used as the output of attention2.

4. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, characterized in that The data processing process of the main classifier specifically includes: Receive the feature map map2_6 output by attention2. After the feature map map2_6 goes through global average pooling and Reshape operations in sequence, a feature vector F2_1 is obtained. After the feature vector F2_1 goes through a fully connected layer FC, a feature vector F2_2 is obtained, and the feature vector F2_2 is used as the output of the main classifier.

5. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, characterized in that, The data processing process of the auxiliary classifier 1 specifically includes: Receive the feature map map3_1 output by layer2 in the ResNeXt101 network. After the feature map map3_1 goes through global average pooling and Reshape operations in sequence, a feature vector F3_1 is obtained. After the feature vector F3_1 goes through a fully connected layer FC, a feature vector F3_2 is obtained. After the feature vector F3_2 is multiplied by a trainable weight vector W1, a feature vector F3_3 is obtained, and the feature vector F3_3 is used as the output of the auxiliary classifier 1.

6. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, characterized in that The data processing process of the auxiliary classifier 2 specifically includes: Receive the feature map map4_1 output by layer3 in the ResNeXt101 network. After the feature map map4_1 goes through global average pooling and Reshape operations in sequence, a feature vector F4_1 is obtained. After the feature vector F4_1 goes through a fully connected layer FC, a feature vector F4_2 is obtained. After the feature vector F4_2 is multiplied by a trainable weight vector W2, a feature vector F4_3 is obtained, and the feature vector F4_3 is used as the output of the auxiliary classifier 2.

7. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, characterized in that In step 1, use the trained object detection model to locate the lumbar spine image to be measured.

8. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, wherein, Step 1 also includes: performing image size normalization processing on the effective lumbar spine region image; Correspondingly, step 4 is: input the effective lumbar spine region image after image size normalization processing into the trained intelligent grading network model for lumbar disc herniation to obtain the grade information of lumbar disc herniation.

9. The intelligent grading method for lumbar disc herniation based on global attention CNN according to claim 1, wherein The lumbar spine image to be measured is a T2-weighted axial lumbar spine MRI image, where T2 represents the transverse relaxation time.

10. An intelligent grading system for lumbar disc herniation based on global attention CNN, characterized in that, Include: An image preprocessing module, which is used to locate the lumbar spine image to be measured to obtain an effective lumbar spine region, and then intercept the effective lumbar spine region image; An intelligent grading network module improves the ResNeXt101 network to construct an intelligent grading network for lumbar disc herniation, specifically including: adding a spatial attention module attention1 at the input end of the ResNeXt101 network, adding a spatial attention module attention2 after layer4 in the ResNeXt101 network, and adding a main classifier after attention2; adding an auxiliary classifier 1 and using the feature map output by layer2 in the ResNeXt101 network as the input of the auxiliary classifier 1, adding an auxiliary classifier 2 and using the feature map output by layer3 in the ResNeXt101 network as the input of the auxiliary classifier 2; the sum of the outputs of the main classifier, the output of the auxiliary classifier 1, and the output of the auxiliary classifier 2 is used as the input of the Softmax classifier, and the Softmax classifier outputs the grade information of lumbar disc herniation; A training module is used to train the intelligent grading network for lumbar disc herniation to obtain an intelligent grading network model for lumbar disc herniation; An intelligent grading module is used to input the effective lumbar region image corresponding to the lumbar image to be measured into the trained intelligent grading network model for lumbar disc herniation to obtain the grade information of lumbar disc herniation.

Images