Scoliosis recognition system based on machine learning

Abstract

The present invention discloses a scoliosis recognition system based on machine learning, which belongs to the field of medical image recognition technology, and includes a data collection module, a data preprocessing module, a spinal image segmentation module, a scoliosis recognition module and a comprehensive report generation module. The present invention adopts an improved mask region convolutional neural network model for spinal image segmentation, introduces a composite connection network, so that the model has better performance on complex images, and the channel attention mechanism can better focus on the key information in the image, suppress irrelevant information in the background, and significantly improve the accuracy of spinal image segmentation; an adaptive weight adjustable enhancement algorithm is adopted for scoliosis recognition, and adjustable parameters are introduced to dynamically adjust the weight update mechanism, which improves the performance of the model on complex data and reduces the influence of difficult-to-classify samples, thereby improving the overall stability of the model, reducing the number of iterations, and improving the overall efficiency of the model, which helps to achieve efficient and accurate scoliosis recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image recognition, and specifically refers to a scoliosis recognition system based on machine learning. Background Art

[0002] The machine learning-based scoliosis recognition system is a system that uses machine learning technology to automatically analyze spine-related medical images and identify spinal conditions. It aims to automatically identify scoliosis, help doctors improve work efficiency and reduce workload.

[0003] However, in the existing scoliosis identification process, there is a lot of irrelevant information mixed in the spinal X-rays, which leads to the technical problem that the image segmentation method has low accuracy in segmentation results when dealing with complex spinal structures; there is a large difference in scoliosis images between different patients, and manual identification of scoliosis is prone to observation errors and takes a long time. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a scoliosis recognition system based on machine learning. In view of the technical problem that in the existing scoliosis recognition process, there is a lot of irrelevant information mixed in the spinal X-ray film, which leads to the low accuracy of the segmentation result when the image segmentation method processes complex spinal structures, the present solution creatively adopts an improved mask region convolutional neural network model for spinal image segmentation, introduces a composite connection network, so that the model has better performance in complex images, and the channel attention mechanism can better focus on the key information in the image, suppress irrelevant information in the background, and significantly improve the accuracy of spinal image segmentation; in view of the technical problem that in the existing scoliosis recognition process, there are large differences in scoliosis images of different patients, and manual identification of scoliosis is prone to observation errors and takes a long time, the present solution creatively adopts an adaptive weight adjustable enhancement algorithm for scoliosis recognition, introduces adjustable parameters, and dynamically adjusts the weight update mechanism, which improves the performance of the model on complex data and reduces the influence of difficult-to-classify samples, thereby improving the overall stability and accuracy of the model, reducing the number of iterations, and improving the overall efficiency of the model, which helps to achieve efficient and accurate scoliosis recognition.

[0005] The technical solution adopted by the present invention is as follows: the scoliosis identification system based on machine learning provided by the present invention includes a data collection module, a data preprocessing module, a spinal image segmentation module, a scoliosis identification module and a comprehensive report generation module;

[0006] The data collection module obtains original spinal image data and spinal condition labels through data collection, and sends the original spinal image data to the data preprocessing module, and sends the spinal condition labels to the scoliosis recognition module;

[0007] The data preprocessing module performs data preprocessing through an image enhancement unit, an image adjustment unit and a normalization unit to obtain standard spine image data, and sends the standard spine image data to a spine image segmentation module;

[0008] The spinal image segmentation module uses an improved mask region convolutional neural network model to perform spinal image segmentation to obtain spinal region segmentation results, and sends the spinal region segmentation results to the scoliosis recognition module and the comprehensive report generation module;

[0009] The scoliosis recognition module adopts an adaptive weight adjustable enhancement algorithm to perform scoliosis recognition, obtain spinal classification information, and send the spinal classification information to the comprehensive report generation module;

[0010] The comprehensive report generation module generates a comprehensive report by combining the spinal region segmentation results and the spinal classification information.

[0011] Furthermore, in the data collection module, the data collection is used to collect the original image data required for scoliosis identification, specifically to collect anteroposterior spinal X-rays from the hospital system to obtain the original spinal image data and spinal condition labels.

[0012] Furthermore, in the data preprocessing module, an image enhancement unit, an image adjustment unit and a normalization unit are provided, including the following contents:

[0013] An image enhancement unit adjusts the contrast, saturation and brightness of the image in the original image data of the spine, and performs image denoising through a median filter to obtain spine enhanced image data;

[0014] An image adjustment unit adjusts the image size in the spine enhanced image data to a uniform size to obtain spine adjusted image data;

[0015] The normalization unit normalizes the image in the spinal column adjustment image data to obtain spinal column standard image data.

[0016] Furthermore, in the spine image segmentation module, an improved mask region convolutional neural network model is used to perform spine image segmentation to obtain a spine region segmentation result;

[0017] The improved mask region convolutional neural network model includes an input layer, a feature extraction backbone network, a feature pyramid network, a region proposal optimization network, a candidate region alignment layer and a detection head;

[0018] The feature extraction backbone network is used to generate rich multi-level feature maps;

[0019] The feature pyramid network is used to process multi-scale features in multi-level feature maps;

[0020] The region proposal optimization network is used to generate candidate regions containing spines;

[0021] The candidate region alignment layer is used to align candidate regions;

[0022] The spine image segmentation module is provided with a model improvement unit, an input layer construction unit, a feature extraction backbone network construction unit, a feature pyramid network construction unit, a region proposal optimization network construction unit, a candidate region alignment layer construction unit, a detection head construction unit, an improved mask region convolutional neural network model construction unit and an image segmentation result generation unit, including the following contents:

[0023] The model improvement unit introduces the composite connection network into the backbone network of the masked region convolutional neural network model, and introduces the channel attention mechanism into the region proposal network of the masked region convolutional neural network model, improves the masked region convolutional neural network model, and obtains an improved masked region convolutional neural network model;

[0024] An input layer construction unit, used for receiving an input image, specifically receiving an image in the spine standard image data as an input image through the input layer;

[0025] The feature extraction backbone network construction unit sets the ResNet-101 network and the composite connection network in the feature extraction backbone network. The deep features of the input image are extracted through the feature extraction backbone network to obtain a multi-level feature map. The calculation formula is:

[0026] ;

[0027] In the formula, Ig res is the ResNet-101 network output feature map, f ResNet-101 (·) is the ResNet-101 network calculation function, Ig input is the input image, Ig cb It is a multi-level feature map, specifically the composite connection network output feature map, f CB-Net (·) is the composite connection network calculation function;

[0028] The feature pyramid network construction unit performs multi-scale fusion of multi-level feature maps through the feature pyramid network to obtain a multi-scale feature map. The calculation formula is:

[0029] ;

[0030] Where Igp is a multi-scale feature map, f fpn (·) is the feature pyramid network calculation function;

[0031] The region proposal optimization network building unit includes the following:

[0032] The channel attention mechanism is used to enhance the key channel information in the multi-scale feature map. Specifically, each channel in the multi-scale feature map is first globally averaged pooled, and then the channel attention weight is learned through the fully connected layer. After that, the multi-scale feature map is weighted using the channel attention weight to generate a channel enhanced feature map. The calculation formula is:

[0033] ;

[0034] In the formula, z c is the global average of the cth channel in the multi-scale feature map, c is the channel index, H is the height of the multi-scale feature map, i is the vertical index of the multi-scale feature map, W is the width of the multi-scale feature map, j is the horizontal index of the multi-scale feature map, is the pixel value of the cth channel at position (i, j) in the multi-scale feature map, w c is the attention weight of the cth channel in the multi-scale feature map, Sig(·) is the S-type activation function, and W one is the first weight of the fully connected layer, ReLU(·) is the ReLU activation function, and W two is the second weight of the fully connected layer, pic c is the eigenvalue of the cth channel in the channel enhancement feature map, Igp c is the eigenvalue of the cth channel in the multi-scale feature map;

[0035] The channel enhanced feature map is processed through the region proposal network to generate candidate regions;

[0036] A candidate region alignment layer construction unit aligns the candidate regions through the candidate region alignment layer to obtain candidate alignment regions;

[0037] The detection head construction unit sets the segmentation, classification and bounding box regression tasks in the detection head, processes the candidate alignment area through the detection head, and obtains the final segmentation result;

[0038] An improved mask region convolutional neural network model building unit is used to build an improved mask region convolutional neural network model through an input layer building unit, a feature extraction backbone network building unit, a feature pyramid network building unit, a region proposal optimization network building unit, a candidate region alignment layer building unit, and a detection head building unit.

[0039] The image segmentation result generating unit performs image segmentation through an improved mask region convolutional neural network model to obtain a spinal region segmentation result.

[0040] Furthermore, in the scoliosis recognition module, based on the spinal region segmentation results, an adaptive weight-adjustable enhancement algorithm is used to perform scoliosis recognition and obtain spinal classification information;

[0041] The scoliosis recognition module is provided with an adaptive enhancement algorithm optimization unit, an adaptive weight adjustable enhancement algorithm construction unit and a scoliosis recognition result generation unit, and includes the following contents:

[0042] The adaptive enhancement algorithm optimization unit is used to enhance the flexibility of the algorithm and improve its adaptability to complex data. Specifically, by introducing adjustable parameters into the adaptive enhancement algorithm, the sample weight update mechanism is adjusted to optimize the adaptive enhancement algorithm, thereby obtaining an adaptive weight adjustable enhancement algorithm.

[0043] The adaptive weight adjustable enhancement algorithm building unit includes the following:

[0044] Initialize the sample weights. Specifically, use the spinal region segmentation results as input samples and set the sample weights equally. The calculation formula is:

[0045] ;

[0046] In the formula, is the initial sample weight, specifically the weight of the kth input sample in the first iteration, k is the input sample index, and N is the number of input samples;

[0047] Train the weak classifier, specifically use the current sample weight to train the weak classifier once, and calculate the classification error. The calculation formula is:

[0048] ;

[0049] In the formula, err m is the classification error of the mth weak classifier, m is the iterative common index of the weak classifier, is the weight of the kth input sample in the mth iteration, I(·) is the indicator function, when When the indicator function takes the value of 1, otherwise it takes the value of 0. m (·) is the weak classifier prediction function, x k is the kth input sample, y k is the true label of the kth input sample, and the true label specifically refers to the spinal condition label;

[0050] Calculate the weight of the weak classifier, specifically, calculate the weight of the weak classifier based on the classification error and adjustable parameters, and the calculation formula is:

[0051] ;

[0052] In the formula, A m is the weight of the mth weak classifier, is an adjustable parameter, log(·) is a logarithmic function;

[0053] Update the sample weights, specifically by controlling the update amplitude of the sample weights through adjustable parameters, and update the sample weights. The calculation formula is:

[0054] ;

[0055] In the formula, is the updated sample weight, specifically the weight of the kth input sample in the m+1th iteration, Z m is the normalization factor, e is a natural constant;

[0056] Constructing a weak classifier, specifically by repeatedly executing the initialization sample weight, the training weak classifier, the calculation of the weak classifier weight and the update sample weight until a maximum number of iterations is reached, thereby constructing the weak classifier;

[0057] Construct a combined classifier, specifically by performing a weighted linear combination of all weak classifiers to obtain a combined classifier. The calculation formula is:

[0058] ;

[0059] Where Q(·) is the combined classifier prediction function, M is the number of weak classifiers, and the number of weak classifiers is equal to the number of iterations;

[0060] The scoliosis recognition result generating unit uses a combined classifier to perform scoliosis recognition and obtain spinal classification information.

[0061] Furthermore, in the comprehensive report generation module, the spinal region segmentation results and spinal classification information are combined to generate a comprehensive scoliosis identification report and provide treatment recommendations.

[0062] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0063] (1) In order to solve the technical problem that spinal X-rays contain a lot of irrelevant information in the existing scoliosis recognition process, resulting in low accuracy of segmentation results when image segmentation methods process complex spinal structures, this solution creatively adopts an improved mask region convolutional neural network model for spinal image segmentation and introduces a composite connection network, which makes the model more capable of performing in complex images. The channel attention mechanism can better focus on the key information in the image and suppress irrelevant information in the background, significantly improving the accuracy of spinal image segmentation.

[0064] (2) In view of the technical problems that the scoliosis images of different patients vary greatly in the existing scoliosis identification process, manual identification of scoliosis is prone to observation errors and takes a long time, this solution creatively adopts an adaptive weight adjustable enhancement algorithm for scoliosis identification, introduces adjustable parameters, and dynamically adjusts the weight update mechanism, which improves the performance of the model on complex data and reduces the impact of difficult-to-classify samples, thereby improving the overall stability and accuracy of the model, reducing the number of iterations, and improving the overall efficiency of the model, which helps to achieve efficient and accurate scoliosis identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A schematic diagram of the process of a scoliosis identification system based on machine learning provided by the present invention;

[0066] Figure 2 It is a flowchart of the data preprocessing module;

[0067] Figure 3 It is a flowchart of the spine image segmentation module;

[0068] Figure 4 Schematic diagram of the process of scoliosis identification module.

[0069] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0071] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0072] Example 1, see Figure 1 , the scoliosis identification system based on machine learning provided by the present invention includes a data collection module, a data preprocessing module, a spinal image segmentation module, a scoliosis identification module and a comprehensive report generation module;

[0073] The data collection module obtains original spinal image data and spinal condition labels through data collection, and sends the original spinal image data to the data preprocessing module, and sends the spinal condition labels to the scoliosis recognition module;

[0074] The data preprocessing module performs data preprocessing through an image enhancement unit, an image adjustment unit and a normalization unit to obtain standard spine image data, and sends the standard spine image data to a spine image segmentation module;

[0075] The spinal image segmentation module uses an improved mask region convolutional neural network model to perform spinal image segmentation to obtain spinal region segmentation results, and sends the spinal region segmentation results to the scoliosis recognition module and the comprehensive report generation module;

[0076] The scoliosis recognition module adopts an adaptive weight adjustable enhancement algorithm to perform scoliosis recognition, obtain spinal classification information, and send the spinal classification information to the comprehensive report generation module;

[0077] The comprehensive report generation module generates a comprehensive report by combining the spinal region segmentation results and the spinal classification information.

[0078] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the data collection module, the data collection is used to collect the original image data required for scoliosis identification, specifically collecting anteroposterior spinal X-rays from the hospital system to obtain the original spinal image data and spinal condition labels;

[0079] The spinal condition labels include normal spine, C-type scoliosis and S-type scoliosis.

[0080] Example 3, see Figure 1 and Figure 2This embodiment is based on the above embodiment. In the data preprocessing module, an image enhancement unit, an image adjustment unit and a normalization unit are provided, and include the following contents:

[0081] An image enhancement unit adjusts the contrast, saturation and brightness of the image in the original image data of the spine, and performs image denoising through a median filter to obtain spine enhanced image data;

[0082] An image adjustment unit adjusts the image size in the spine enhanced image data to a uniform size to obtain spine adjusted image data;

[0083] The normalization unit normalizes the image in the spinal column adjustment image data to obtain spinal column standard image data.

[0084] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the spine image segmentation module, an improved mask region convolutional neural network model is used to perform spine image segmentation to obtain a spine region segmentation result;

[0085] The improved mask region convolutional neural network model includes an input layer, a feature extraction backbone network, a feature pyramid network, a region proposal optimization network, a candidate region alignment layer and a detection head;

[0086] The feature extraction backbone network is used to generate rich multi-level feature maps;

[0087] The feature pyramid network is used to process multi-scale features in multi-level feature maps;

[0088] The region proposal optimization network is used to generate candidate regions containing spines;

[0089] The candidate region alignment layer is used to align candidate regions;

[0090] The spine image segmentation module is provided with a model improvement unit, an input layer construction unit, a feature extraction backbone network construction unit, a feature pyramid network construction unit, a region proposal optimization network construction unit, a candidate region alignment layer construction unit, a detection head construction unit, an improved mask region convolutional neural network model construction unit and an image segmentation result generation unit, including the following contents:

[0091] The model improvement unit introduces the composite connection network into the backbone network of the masked region convolutional neural network model, and introduces the channel attention mechanism into the region proposal network of the masked region convolutional neural network model, improves the masked region convolutional neural network model, and obtains an improved masked region convolutional neural network model;

[0092] An input layer construction unit, used for receiving an input image, specifically receiving an image in the spine standard image data as an input image through the input layer;

[0093] The feature extraction backbone network construction unit sets the ResNet-101 network and the composite connection network in the feature extraction backbone network. The deep features of the input image are extracted through the feature extraction backbone network to obtain a multi-level feature map. The calculation formula is:

[0094] ;

[0095] In the formula, Ig res is the ResNet-101 network output feature map, f ResNet-101 (·) is the ResNet-101 network calculation function, Ig input is the input image, Ig cb It is a multi-level feature map, specifically the composite connection network output feature map, f CB-Net (·) is the composite connection network calculation function;

[0096] The feature pyramid network construction unit performs multi-scale fusion of multi-level feature maps through the feature pyramid network to obtain a multi-scale feature map. The calculation formula is:

[0097] ;

[0098] Where Igp is a multi-scale feature map, f fpn (·) is the feature pyramid network calculation function;

[0099] The region proposal optimization network building unit includes the following:

[0100] The channel attention mechanism is used to enhance the key channel information in the multi-scale feature map. Specifically, each channel in the multi-scale feature map is first globally averaged pooled, and then the channel attention weight is learned through the fully connected layer. After that, the multi-scale feature map is weighted using the channel attention weight to generate a channel enhanced feature map. The calculation formula is:

[0101] ;

[0102] In the formula, z c is the global average of the cth channel in the multi-scale feature map, c is the channel index, H is the height of the multi-scale feature map, i is the vertical index of the multi-scale feature map, W is the width of the multi-scale feature map, j is the horizontal index of the multi-scale feature map, is the pixel value of the cth channel at position (i, j) in the multi-scale feature map, w c is the attention weight of the cth channel in the multi-scale feature map, Sig(·) is the S-type activation function, and Wone is the first weight of the fully connected layer, ReLU(·) is the ReLU activation function, and W two is the second weight of the fully connected layer, pic c is the eigenvalue of the cth channel in the channel enhancement feature map, Igp c is the eigenvalue of the cth channel in the multi-scale feature map;

[0103] The channel enhanced feature map is processed through the region proposal network to generate candidate regions;

[0104] A candidate region alignment layer construction unit aligns the candidate regions through the candidate region alignment layer to obtain candidate alignment regions;

[0105] The detection head construction unit sets the segmentation, classification and bounding box regression tasks in the detection head, processes the candidate alignment area through the detection head, and obtains the final segmentation result;

[0106] An improved mask region convolutional neural network model building unit is used to build an improved mask region convolutional neural network model through an input layer building unit, a feature extraction backbone network building unit, a feature pyramid network building unit, a region proposal optimization network building unit, a candidate region alignment layer building unit, and a detection head building unit.

[0107] An image segmentation result generating unit performs image segmentation through an improved mask region convolutional neural network model to obtain a spinal region segmentation result;

[0108] The spine region segmentation result includes a spine region segmentation mask, a bounding box, and a category label;

[0109] The category labels include normal morphology and abnormal morphology;

[0110] Table 1 is a comparison of model performance before and after the mask region convolutional neural network model is improved in Example 4 of the present invention. As shown in the table, FPS is used to represent the number of images processed per second, and the higher the FPS, the faster the model processing speed;

[0111] Table 1 Comparison of model performance before and after improvement of the mask region convolutional neural network model

[0112]

[0113] By performing the above operations, in order to solve the technical problem that in the existing scoliosis identification process, there is a lot of irrelevant information mixed in the spinal X-rays, which leads to low accuracy of segmentation results when the image segmentation method processes complex spinal structures, this scheme creatively adopts an improved mask region convolutional neural network model for spinal image segmentation and introduces a composite connection network to make the model perform better in complex images. The channel attention mechanism can better focus on the key information in the image and suppress irrelevant information in the background, which significantly improves the accuracy of spinal image segmentation.

[0114] Example 5, see Figure 1 and Figure 4 , based on the above embodiment, in the scoliosis recognition module, according to the spinal region segmentation result, an adaptive weight adjustable enhancement algorithm is used to perform scoliosis recognition and obtain spinal classification information;

[0115] The scoliosis recognition module is provided with an adaptive enhancement algorithm optimization unit, an adaptive weight adjustable enhancement algorithm construction unit and a scoliosis recognition result generation unit, and includes the following contents:

[0116] The adaptive enhancement algorithm optimization unit is used to enhance the flexibility of the algorithm and improve its adaptability to complex data. Specifically, by introducing adjustable parameters into the adaptive enhancement algorithm, the sample weight update mechanism is adjusted to optimize the adaptive enhancement algorithm, thereby obtaining an adaptive weight adjustable enhancement algorithm.

[0117] The adaptive weight adjustable enhancement algorithm building unit includes the following:

[0118] Initialize the sample weights. Specifically, use the spinal region segmentation results as input samples and set the sample weights equally. The calculation formula is:

[0119] ;

[0120] In the formula, is the initial sample weight, specifically the weight of the kth input sample in the first iteration, k is the input sample index, and N is the number of input samples;

[0121] Train the weak classifier, specifically use the current sample weight to train the weak classifier once, and calculate the classification error. The calculation formula is:

[0122] ;

[0123] In the formula, err m is the classification error of the mth weak classifier, m is the iterative common index of the weak classifier, is the weight of the kth input sample in the mth iteration, I(·) is the indicator function, when When the indicator function takes the value of 1, otherwise it takes the value of 0. m (·) is the weak classifier prediction function, x k is the kth input sample, y k is the true label of the kth input sample, and the true label specifically refers to the spinal condition label;

[0124] Calculate the weight of the weak classifier, specifically, calculate the weight of the weak classifier based on the classification error and adjustable parameters, and the calculation formula is:

[0125] ;

[0126] In the formula, A m is the weight of the mth weak classifier, is an adjustable parameter, log(·) is a logarithmic function;

[0127] Update the sample weights, specifically by controlling the update amplitude of the sample weights through adjustable parameters, and update the sample weights. The calculation formula is:

[0128] ;

[0129] In the formula, is the updated sample weight, specifically the weight of the kth input sample in the m+1th iteration, Z m is the normalization factor, e is a natural constant;

[0130] Constructing a weak classifier, specifically by repeatedly executing the initialization sample weight, the training weak classifier, the calculation of the weak classifier weight and the update sample weight until a maximum number of iterations is reached, thereby constructing the weak classifier;

[0131] Construct a combined classifier, specifically by performing a weighted linear combination of all weak classifiers to obtain a combined classifier. The calculation formula is:

[0132] ;

[0133] Where Q(·) is the combined classifier prediction function, M is the number of weak classifiers, and the number of weak classifiers is equal to the number of iterations;

[0134] A scoliosis recognition result generating unit uses a combined classifier to perform scoliosis recognition and obtain spinal classification information;

[0135] By performing the above operations, in order to address the technical problems that in the existing scoliosis recognition process, there are large differences in scoliosis images of different patients, manual identification of scoliosis is prone to observation errors, and it takes a long time, this solution creatively adopts an adaptive weight adjustable enhancement algorithm for scoliosis recognition, introduces adjustable parameters, and dynamically adjusts the weight update mechanism, which improves the performance of the model on complex data and reduces the impact of difficult-to-classify samples, thereby improving the overall stability and accuracy of the model, reducing the number of iterations, and improving the overall efficiency of the model, which helps to achieve efficient and accurate scoliosis recognition.

[0136] Example 6, see Figure 1 ,This embodiment is based on the above embodiment, in a comprehensive report generation module, combines the spinal region segmentation results and spinal classification information, generates a scoliosis identification comprehensive report, and provides treatment suggestions.

[0137] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0138] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

[0139] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A scoliosis recognition system based on machine learning, characterized by: It includes a data collection module, a data preprocessing module, a spinal image segmentation module, a scoliosis recognition module and a comprehensive report generation module; The data collection module obtains original spinal image data and spinal condition labels through data collection, and sends the original spinal image data to the data preprocessing module, and sends the spinal condition labels to the scoliosis recognition module; The data preprocessing module performs data preprocessing through an image enhancement unit, an image adjustment unit and a normalization unit to obtain standard spine image data, and sends the standard spine image data to a spine image segmentation module; The spinal image segmentation module uses an improved mask region convolutional neural network model to perform spinal image segmentation to obtain spinal region segmentation results, and sends the spinal region segmentation results to the scoliosis recognition module and the comprehensive report generation module; The improved mask region convolutional neural network model includes an input layer, a feature extraction backbone network, a feature pyramid network, a region proposal optimization network, a candidate region alignment layer and a detection head; The feature extraction backbone network is used to generate rich multi-level feature maps; The feature pyramid network is used to process multi-scale features in multi-level feature maps; The region proposal optimization network is used to generate candidate regions containing spines; The candidate region alignment layer is used to align candidate regions; The spine image segmentation module is provided with a model improvement unit, an input layer construction unit, a feature extraction backbone network construction unit, a feature pyramid network construction unit, a region proposal optimization network construction unit, a candidate region alignment layer construction unit, a detection head construction unit, an improved mask region convolutional neural network model construction unit and an image segmentation result generation unit, including the following contents: The model improvement unit introduces the composite connection network into the backbone network of the masked region convolutional neural network model, and introduces the channel attention mechanism into the region proposal network of the masked region convolutional neural network model, improves the masked region convolutional neural network model, and obtains an improved masked region convolutional neural network model; An input layer construction unit, used for receiving an input image, specifically receiving an image in the spine standard image data as an input image through the input layer; The feature extraction backbone network construction unit sets the ResNet-101 network and the composite connection network in the feature extraction backbone network. The deep features of the input image are extracted through the feature extraction backbone network to obtain a multi-level feature map. The calculation formula is: ； In the formula, Ig res is the ResNet-101 network output feature map, f ResNet-101 (·) is the ResNet-101 network calculation function, Ig input is the input image, Ig cb It is a multi-level feature map, specifically the composite connection network output feature map, f CB-Net (·) is the composite connection network calculation function; The feature pyramid network construction unit performs multi-scale fusion of multi-level feature maps through the feature pyramid network to obtain a multi-scale feature map. The calculation formula is: ； Where Igp is a multi-scale feature map, f fpn (·) is the feature pyramid network calculation function; The region proposal optimization network building unit includes the following: The channel attention mechanism is used to enhance the key channel information in the multi-scale feature map. Specifically, each channel in the multi-scale feature map is first globally averaged pooled, and then the channel attention weight is learned through the fully connected layer. After that, the multi-scale feature map is weighted using the channel attention weight to generate a channel enhanced feature map. The calculation formula is: ； In the formula, z c is the global average of the cth channel in the multi-scale feature map, c is the channel index, H is the height of the multi-scale feature map, i is the vertical index of the multi-scale feature map, W is the width of the multi-scale feature map, j is the horizontal index of the multi-scale feature map, is the pixel value of the cth channel at position (i, j) in the multi-scale feature map, w c is the attention weight of the cth channel in the multi-scale feature map, Sig(·) is the S-type activation function, and W one is the first weight of the fully connected layer, ReLU(·) is the ReLU activation function, and W two is the second weight of the fully connected layer, pic c is the eigenvalue of the cth channel in the channel enhancement feature map, Igp c is the eigenvalue of the cth channel in the multi-scale feature map; The channel enhanced feature map is processed through the region proposal network to generate candidate regions; A candidate region alignment layer construction unit aligns the candidate regions through the candidate region alignment layer to obtain candidate alignment regions; The detection head construction unit sets the segmentation, classification and bounding box regression tasks in the detection head, processes the candidate alignment area through the detection head, and obtains the final segmentation result; An improved mask region convolutional neural network model building unit is used to build an improved mask region convolutional neural network model through an input layer building unit, a feature extraction backbone network building unit, a feature pyramid network building unit, a region proposal optimization network building unit, a candidate region alignment layer building unit, and a detection head building unit. An image segmentation result generating unit performs image segmentation through an improved mask region convolutional neural network model to obtain a spinal region segmentation result; The scoliosis recognition module uses an adaptive weight adjustable enhancement algorithm to perform scoliosis recognition based on the spinal region segmentation result, obtains spinal classification information, and sends the spinal classification information to the comprehensive report generation module; The scoliosis recognition module is provided with an adaptive enhancement algorithm optimization unit, an adaptive weight adjustable enhancement algorithm construction unit and a scoliosis recognition result generation unit, and includes the following contents: The adaptive enhancement algorithm optimization unit adjusts the sample weight update mechanism by introducing adjustable parameters into the adaptive enhancement algorithm, optimizes the adaptive enhancement algorithm, and obtains an adaptive weight adjustable enhancement algorithm; The adaptive weight adjustable enhancement algorithm building unit includes the following: Initialize the sample weights. Specifically, use the spinal region segmentation results as input samples and set the sample weights equally. The calculation formula is: ； In the formula, is the initial sample weight, specifically the weight of the kth input sample in the first iteration, k is the input sample index, and N is the number of input samples; Train the weak classifier, specifically use the current sample weight to train the weak classifier once, and calculate the classification error. The calculation formula is: ； In the formula, err m is the classification error of the mth weak classifier, m is the iterative common index of the weak classifier, is the weight of the kth input sample in the mth iteration, I(·) is the indicator function, when When the indicator function takes the value of 1, otherwise it takes the value of 0. m (·) is the weak classifier prediction function, x k is the kth input sample, y k is the true label of the kth input sample, and the true label specifically refers to the spinal condition label; Calculate the weight of the weak classifier, specifically, calculate the weight of the weak classifier based on the classification error and adjustable parameters, and the calculation formula is: ； In the formula, A m is the weight of the mth weak classifier, is an adjustable parameter, log(·) is a logarithmic function; Update the sample weights, specifically by controlling the update amplitude of the sample weights through adjustable parameters, and update the sample weights. The calculation formula is: ； In the formula, is the updated sample weight, specifically the weight of the kth input sample in the m+1th iteration, Z m is the normalization factor, e is a natural constant; Constructing a weak classifier, specifically by repeatedly executing the initialization sample weight, the training weak classifier, the calculation of the weak classifier weight and the update sample weight until a maximum number of iterations is reached, thereby constructing the weak classifier; Construct a combined classifier, specifically by performing a weighted linear combination of all weak classifiers to obtain a combined classifier. The calculation formula is: ； Where Q(·) is the combined classifier prediction function, M is the number of weak classifiers, and the number of weak classifiers is equal to the number of iterations; A scoliosis recognition result generating unit uses a combined classifier to perform scoliosis recognition and obtain spinal classification information; The comprehensive report generation module generates a comprehensive report by combining the spinal region segmentation results and the spinal classification information.

2. The scoliosis identification system based on machine learning according to claim 1, characterized in that: In the data preprocessing module, there are an image enhancement unit, an image adjustment unit and a normalization unit, including the following contents: An image enhancement unit adjusts the contrast, saturation and brightness of the image in the original image data of the spine, and performs image denoising through a median filter to obtain spine enhanced image data; An image adjustment unit adjusts the image size in the spine enhanced image data to a uniform size to obtain spine adjusted image data; The normalization unit normalizes the images in the spinal column adjustment image data to obtain spinal column standard image data.

3. The scoliosis identification system based on machine learning according to claim 2, characterized in that: In the data collection module, the data collection is used to collect the original image data required for scoliosis identification, specifically collecting anteroposterior spinal X-rays from the hospital system to obtain the original spinal image data and spinal condition labels.

4. The scoliosis identification system based on machine learning according to claim 3, characterized in that: In the comprehensive report generation module, the spinal region segmentation results and spinal classification information are combined to generate a comprehensive scoliosis identification report and provide treatment recommendations.

Images