Esophageal cancer nuclear magnetic diagnosis enhancing system based on artificial intelligence

Abstract

The invention discloses an esophageal cancer nuclear magnetic diagnosis enhancement system based on artificial intelligence, which belongs to the field of image processing and comprises an image acquisition module, a nuclear magnetic image optimization module, an esophageal cancer nuclear magnetic diagnosis model building module and a diagnosis enhancement module. The method specifically comprises the steps of performing multi-scale wavelet decomposition, designing an initial position of chaotic sine correlation mapping, constructing two neighborhoods based on a historical optimal position, completing updating of an individual position in combination with cosine similarity segmentation, finding an optimal threshold value, performing threshold value processing on a detail coefficient, and performing wavelet reconstruction to obtain an optimized nuclear magnetic image. Carrying out primary feature extraction to obtain a multi-scale feature map, carrying out two-time different aggregation on advanced features to extract global features, extracting local features from the first feature map, carrying out selective fusion on the global features and the local features to obtain final features, and outputting a diagnosis result to improve the quality of the nuclear magnetic image. And the accuracy and reliability of esophageal cancer diagnosis are further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and specifically refers to an esophageal cancer nuclear magnetic resonance diagnosis enhancement system based on artificial intelligence. Background Art

[0002] Magnetic resonance imaging (MRI) diagnosis of esophageal cancer uses advanced artificial intelligence technology to conduct in-depth analysis and processing of MRI images of esophageal cancer patients, which can detect signs of esophageal cancer more quickly and objectively and reduce human errors. However, the existing MRI diagnosis system for esophageal cancer has various noises generated during the MRI imaging process, resulting in blurring, poor contrast and missing details in the acquired MRI images of esophageal cancer, which reduces the accuracy of esophageal cancer diagnosis; the existing MRI diagnosis system for esophageal cancer has the problem that MRI images of esophageal cancer contain a lot of detailed information and semantic information, and a single feature is difficult to fully describe the lesion information. At the same time, the diagnosis of esophageal cancer needs to consider local lesions and the overall condition, and lacks the comprehensive use of global and local information, making it impossible to accurately diagnose esophageal cancer. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an esophageal cancer nuclear magnetic resonance diagnosis enhancement system based on artificial intelligence. In view of the problem that various noises are generated during the nuclear magnetic resonance imaging process in the existing esophageal cancer nuclear magnetic resonance diagnosis system, resulting in blurring, poor contrast and missing details in the acquired esophageal cancer nuclear magnetic resonance images, and reducing the diagnostic accuracy of esophageal cancer, this scheme performs multi-scale wavelet decomposition on the nuclear magnetic resonance image to more accurately identify effective nuclear magnetic resonance image information; designs the initial position of the chaotic sine correlation mapping so that the initial position setting method is random but has a certain regularity; constructs two neighborhoods based on the historical optimal position, and completes the update of the individual position in segments according to the reference neighbors in the neighborhood and combined with the cosine similarity to find the optimal threshold, and determines the position update direction by referring to the neighbors and similarity, thereby accelerating the process of finding the optimal threshold; threshold processing is performed on the detail coefficient based on the optimal threshold, and then wavelet reconstruction is performed to obtain the optimized nuclear magnetic resonance image, thereby improving the quality of the nuclear magnetic resonance image. , further improving the accuracy and reliability of esophageal cancer diagnosis; in view of the existing esophageal cancer MRI diagnosis system, the esophageal cancer MRI images contain a lot of detailed information and semantic information, and a single feature is difficult to fully describe the lesion information. At the same time, the diagnosis of esophageal cancer needs to consider local lesions and the overall condition. There is a lack of comprehensive utilization of global and local information, and it is impossible to accurately diagnose esophageal cancer. This scheme performs preliminary feature extraction on the optimized MRI images to obtain multi-scale feature maps, and obtains esophageal cancer features at different levels and scales; performs two different aggregations on high-level features to extract global features, and more comprehensively captures global feature patterns and structural information; extracts local features from the first feature map, selectively fuses global features and local features, obtains two selectively fused feature maps, connects and processes them to obtain the final features, and outputs the diagnosis results, effectively combining the overall grasp of the global situation with the fine local information, and improving the comprehensiveness and accuracy of esophageal cancer MRI diagnosis.

[0004] The present invention provides an artificial intelligence-based esophageal cancer nuclear magnetic resonance diagnosis enhancement system, which includes an image acquisition module, a nuclear magnetic resonance image optimization module, a module for building an esophageal cancer nuclear magnetic resonance diagnosis model, and a diagnosis enhancement module;

[0005] The image acquisition module acquires historical MRI image data of patients with esophageal cancer;

[0006] The NMR image optimization module performs multi-scale wavelet decomposition on the NMR image, designs the initial position of the chaotic sine correlation mapping, constructs two neighborhoods based on the historical optimal position, updates the individual position in segments based on the reference neighbors in the neighborhood and in combination with the cosine similarity, finds the optimal threshold, performs threshold processing on the detail coefficient based on the optimal threshold, and then performs wavelet reconstruction to obtain the optimized NMR image;

[0007] The module for constructing the esophageal cancer nuclear magnetic resonance diagnosis model performs preliminary feature extraction on the optimized nuclear magnetic resonance image to obtain a multi-scale feature map, performs two different aggregations on the high-level features to extract global features, extracts local features from the first feature map, selectively fuses the global features and the local features to obtain two selectively fused feature maps, connects and processes them to obtain the final features, and outputs the diagnosis result;

[0008] The diagnosis enhancement module assists in understanding the current status of the patient to be diagnosed with esophageal cancer based on the diagnosis results output by the esophageal cancer nuclear magnetic resonance diagnosis model.

[0009] Furthermore, the image acquisition module acquires historical esophageal cancer patient MRI image data; the historical esophageal cancer patient MRI image data includes MRI images and diagnosis results; the diagnosis results include normal, early, middle and late stages, and the diagnosis results are used as data labels.

[0010] Furthermore, the nuclear magnetic resonance image optimization module is provided with an image decomposition unit, an optimal threshold search unit, a threshold processing unit and an image reconstruction unit, and specifically includes the following contents:

[0011] The image decomposition unit uses db8 as the wavelet basis function to perform multi-scale wavelet decomposition on the nuclear magnetic image to obtain approximate coefficients and detail coefficients of multiple scales. The approximate coefficient is the low-frequency information of the nuclear magnetic image, and the detail coefficient is the high-frequency information of the nuclear magnetic image in the horizontal, vertical and diagonal directions;

[0012] The optimal threshold search unit performs threshold processing on the detail coefficients of each scale and each direction, searches for the threshold based on the swarm intelligence algorithm, and finds the optimal threshold for each scale and each direction, including the following:

[0013] Initial position: establish a search space based on the threshold of each scale and each direction, and randomly initialize the first individual position in the search space , design chaotic sinusoidal correlation mapping to initialize the remaining 2-I individual positions, the individual position dimension is equal to the number of scales × the number of directions, use the individual position as the representative of the threshold, and use the peak signal-to-noise ratio of the nuclear magnetic resonance image after threshold processing as the fitness value of the corresponding individual position; the formula used is as follows:

[0014] ;

[0015] In the formula, and are the initial positions of the i+1th and ith individuals in the jth dimension, i is the individual index, j is the dimension index, and r 1 is the first random number, r 2 is the second random number, sin(•) is the sine function;

[0016] Build a neighborhood and calculate the location of each individual and The Euclidean distance based on the closest The k individual positions of construct the first neighborhood of the i-th individual at the t-th iteration , based on the closest The k individual positions of construct the second neighborhood at the tth iteration ; where t is the iteration index, is the historical optimal position of the ith individual at the tth iteration, is the historical optimal position of individual I at the tth iteration. The historical optimal position is the position of the individual with the highest fitness value since the iteration;

[0017] Individual location updates, from and Randomly select the first reference neighbor and the second reference neighbor ,calculate and The cosine similarity between , based on the cosine similarity segmentation, the individual position is updated, and the formula used is as follows:

[0018] ;

[0019] In the formula, and are the positions of the i-th individual in the j-th dimension at the t+1-th and t-th iterations, respectively. and They are and At the position of the jth dimension, and They are and At the position in the jth dimension, r 3 and r 4 are the third and fourth random numbers respectively, W t is the inertia weight at the tth iteration, and are the first acceleration coefficient and the second acceleration coefficient at the tth iteration, is the L2 norm;

[0020] The optimal threshold is determined, the fitness threshold and the maximum number of iterations are pre-set, and the fitness value and ,when When the corresponding fitness value is higher than the fitness threshold, The threshold represented by is taken as the optimal threshold; otherwise, if the maximum number of iterations is reached, the number of iterations is set to 0 and the initial position is reset; otherwise, the number of iterations is increased by 1 and the neighborhood and individual position are rebuilt and updated, and the iteration continues;

[0021] The threshold processing unit designs a threshold processing function and performs threshold processing on the detail coefficients of each scale and each direction based on the optimal threshold. The formula used is as follows:

[0022] ;

[0023] In the formula, and are the original detail coefficients in the fth direction of the cth scale and the detail coefficients after threshold processing, c and f are the scale index and direction index respectively. is the optimal threshold in the fth direction at the cth scale, b is the adjustment coefficient, and sgn (•) is the sign function;

[0024] The image reconstruction unit performs wavelet reconstruction on the original approximate coefficients and the detail coefficients after processing based on the optimal threshold value to obtain an optimized nuclear magnetic resonance image.

[0025] Furthermore, the module for constructing the esophageal cancer nuclear magnetic resonance diagnostic model is provided with a preliminary feature extraction unit, a global feature extraction unit, a local feature extraction unit, a feature fusion unit and an output unit, and specifically includes the following contents:

[0026] The preliminary feature extraction unit inputs the optimized nuclear magnetic resonance image of size A×B into the pyramid visual transformer PVTv2-B2 for processing to obtain a multi-scale feature map, specifically a The first characteristic graph E 1 , size is The second characteristic graph E 2 , size is The third characteristic graph E 3 and size The fourth characteristic graph E 4 , the first feature map E 1 is a low-level feature, the second feature map E 2 , the third characteristic graph E 3 And the fourth characteristic graph E 4 is a high-level feature; where A and B are the image height and width respectively;

[0027] Extract global feature units, perform two different aggregations on high-level features, and calculate the global features of the MRI image based on the two aggregation results, including the following:

[0028] The first aggregation, for the third feature map E 3 And the fourth characteristic graph E4 To perform aggregation, the formula used is as follows:

[0029] ;

[0030] In the formula, R 1 It is the high-level feature of the first aggregation. 2 (•) is a 2x nearest neighbor upsampling operation, is the element-wise product operator, Concat (•) is the concatenation operation, and LE (•) is a series of operations consisting of a 3×3 convolutional layer, a GN normalization layer, and a SiLU activation function;

[0031] The second aggregation, for the second feature map E 2 , the third characteristic graph E 3 And the fourth characteristic graph E 4 To perform aggregation, the formula used is as follows:

[0032] ;

[0033] In the formula, R 2 It is the advanced feature of the second aggregation, Up 4 (•) is a 4x nearest neighbor upsampling operation;

[0034] Based on R 1 and R 2 Get the global feature map, the formula used is as follows:

[0035] ;

[0036] In the formula, Q 1 is the global feature map;

[0037] Extract local feature units from the first feature map E 1 The local features of the nuclear magnetic resonance image are extracted from the image. The formula used is as follows:

[0038] ;

[0039] In the formula, Q 2 is the local feature map, C 1×1 (•) is a 1×1 convolution operation, SiLU (•) is a SiLU activation function, and GN (•) is a normalization operation;

[0040] The feature fusion unit selectively fuses global features and local features to obtain the final features of the MRI image, including the following:

[0041] Selective fusion, the formula used is as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] Where σ(•) is the Sigmoid activation function, and minus D 2 and D 1 The reverse operation, is the element-wise addition operator, D 1 and D 2 They are the processed global feature map and local feature map, and Selective fusion features Figure 1 and selective fusion features Figure 2 ;

[0047] right and Connect to get the connection characteristics , a series of operations consisting of two 1×1 convolution layers, GN normalization layers, and SiLU activation functions are performed on the connection features to obtain the final features of the nuclear magnetic resonance image. The formula used is as follows:

[0048] ;

[0049] Where G is the final feature;

[0050] The output unit processes the final feature G through the fully connected layer and the Softmax function to obtain the probability vector Y, and selects the data label with the highest probability as the final output diagnosis result. The formula used is as follows:

[0051] ;

[0052] Where FC (•) is the fully connected layer and Softmax (•) is the normalized exponential function.

[0053] Furthermore, the diagnosis enhancement module collects MRI images of patients to be diagnosed with esophageal cancer, optimizes the MRI images, and then inputs them into the esophageal cancer MRI diagnosis model for processing, and assists in understanding the current status of patients to be diagnosed with esophageal cancer based on the diagnosis results output by the model.

[0054] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0055] (1) In view of the problem that various noises are generated during the MRI imaging process in the existing esophageal cancer MRI diagnosis system, resulting in blurring, poor contrast and missing details in the acquired esophageal cancer MRI images, which reduces the diagnostic accuracy of esophageal cancer, this scheme performs multi-scale wavelet decomposition on the MRI images, which helps to more accurately identify noise and effective MRI image information; designs the initial position of the chaotic sine correlation mapping so that the initial position setting method is random but has a certain regularity, avoiding blindness; constructs two neighborhoods based on the historical optimal position, and completes the update of individual positions in segments based on the reference neighbors in the neighborhood and the cosine similarity to find the optimal threshold. The position update direction is determined by the reference neighbors and similarity, which accelerates the process of finding the optimal threshold; threshold processing is performed on the detail coefficient based on the optimal threshold, and then wavelet reconstruction is performed to obtain the optimized MRI image, which removes noise more accurately, while retaining the effective details of the MRI image to the greatest extent, improves the quality of the MRI image, and further improves the accuracy and reliability of esophageal cancer diagnosis.

[0056] (2) In view of the problem that the existing esophageal cancer MRI diagnosis system contains a large amount of detailed information and semantic information, a single feature is difficult to fully describe the lesion information. At the same time, the diagnosis of esophageal cancer needs to consider local lesions and the overall condition. There is a lack of comprehensive utilization of global and local information, and it is impossible to accurately diagnose esophageal cancer. This scheme performs preliminary feature extraction on the optimized MRI image to obtain a multi-scale feature map, obtains esophageal cancer features at different levels and scales, and contains richer details and semantic information; performs two different aggregations on the high-level features to extract global features, and captures the global feature patterns and structural information more comprehensively through different aggregation methods, which helps to grasp the esophageal cancer condition from a holistic perspective; extracts local features from the first feature map to highlight local subtle features and abnormalities; selectively fuses the global features and local features to obtain two selectively fused feature maps, connects and processes them to obtain the final features, and outputs the diagnosis results, which effectively combines the global overall grasp and the local fine information, so that the diagnosis results can not only reflect the overall condition of esophageal cancer but also pay attention to specific local details, thereby improving the comprehensiveness and accuracy of esophageal cancer MRI diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of an artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer provided by the present invention;

[0058] Figure 2 Schematic diagram of the module for optimizing nuclear magnetic resonance images;

[0059] Figure 3 Schematic diagram of constructing the esophageal cancer MRI diagnostic model module.

[0060] 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

[0061] 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.

[0062] 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.

[0063] Example 1, see Figure 1 The present invention provides an artificial intelligence-based esophageal cancer nuclear magnetic resonance diagnosis enhancement system, which includes an image acquisition module, a nuclear magnetic resonance image optimization module, a module for building an esophageal cancer nuclear magnetic resonance diagnosis model, and a diagnosis enhancement module;

[0064] The image acquisition module acquires historical MRI image data of esophageal cancer patients and sends the data to the MRI image optimization module;

[0065] The nuclear magnetic image optimization module receives the data sent by the image acquisition module, performs multi-scale wavelet decomposition on the nuclear magnetic image, designs the initial position of the chaotic sine correlation map, constructs two neighborhoods based on the historical optimal position, updates the individual position in segments based on the reference neighbors in the neighborhood and in combination with the cosine similarity, finds the optimal threshold, performs threshold processing on the detail coefficient based on the optimal threshold, performs wavelet reconstruction to obtain the optimized nuclear magnetic image, and sends the data to the module for constructing the esophageal cancer nuclear magnetic diagnosis model;

[0066] The module for constructing the esophageal cancer nuclear magnetic resonance diagnostic model receives the data sent by the nuclear magnetic resonance image optimization module, performs preliminary feature extraction on the optimized nuclear magnetic resonance image to obtain a multi-scale feature map, performs two different aggregations on the high-level features to extract global features, extracts local features from the first feature map, selectively fuses the global features and the local features to obtain two selectively fused feature maps, connects and processes them to obtain the final features, outputs the diagnosis results, and sends the data to the diagnosis enhancement module;

[0067] The diagnosis enhancement module receives data sent by the module for constructing the esophageal cancer nuclear magnetic resonance diagnosis model, and assists in understanding the current status of the esophageal cancer patient to be diagnosed based on the diagnosis result output by the esophageal cancer nuclear magnetic resonance diagnosis model.

[0068] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the image acquisition module, the historical esophageal cancer patient MRI image data includes MRI images and diagnosis results; the diagnosis results include normal, early, middle and late stages, and the diagnosis results are used as data labels.

[0069] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the nuclear magnetic resonance image optimization module, an image decomposition unit, an optimal threshold search unit, a threshold processing unit and an image reconstruction unit are provided. Specifically, the following contents are included:

[0070] The image decomposition unit uses db8 as the wavelet basis function and the Mallat algorithm to perform multi-scale wavelet decomposition on the nuclear magnetic image to obtain approximate coefficients and detail coefficients of multiple scales. The approximate coefficient is the low-frequency information of the nuclear magnetic image, and the detail coefficient is the high-frequency information of the nuclear magnetic image in the horizontal, vertical and diagonal directions. The number of scales is 4;

[0071] The optimal threshold search unit performs threshold processing on the detail coefficients of each scale and each direction, searches for the threshold based on the swarm intelligence algorithm, and finds the optimal threshold for each scale and each direction, including the following:

[0072] Initial position: establish a search space based on the threshold of each scale and each direction, and randomly initialize the first individual position in the search space , design chaotic sinusoidal correlation mapping to initialize the remaining 2-I individual positions, the individual position dimension is equal to the number of scales × the number of directions, use the individual position as the representative of the threshold, and use the peak signal-to-noise ratio of the nuclear magnetic resonance image after threshold processing as the fitness value of the corresponding individual position; the formula used is as follows:

[0073] ;

[0074] In the formula, and are the initial positions of the i+1th and ith individuals in the jth dimension, i is the individual index, j is the dimension index, and r 1 is the first random number in the range (0, 1), r 2 is the second random number in the range of (0, 4), sin(•) is the sine function;

[0075] Build a neighborhood and calculate the location of each individual and The Euclidean distance based on the closest The k individual positions of construct the first neighborhood of the i-th individual at the t-th iteration , based on the closest The k individual positions of construct the second neighborhood at the tth iteration ; where t is the iteration index, is the historical optimal position of the ith individual at the tth iteration, is the historical optimal position of individual I at the tth iteration. The historical optimal position is the position of the individual with the highest fitness value since the iteration;

[0076] Individual location updates, from and Randomly select the first reference neighbor and the second reference neighbor ,calculate and The cosine similarity between , based on the cosine similarity segmentation, the individual position is updated, and the formula used is as follows:

[0077] ;

[0078] In the formula, and are the positions of the i-th individual in the j-th dimension at the t+1-th and t-th iterations, respectively. and They are and At the position of the jth dimension, and They are and At the position in the jth dimension, r 3 and r 4 are the third and fourth random numbers in the range of (0, 1), respectively, W t is the inertia weight at the tth iteration, , W max is the maximum inertia weight, exp(•) is the exponential function, and are the first acceleration coefficient and the second acceleration coefficient at the tth iteration, , , T is the maximum number of iterations, cos(•) is the cosine function, is the L2 norm;

[0079] The optimal threshold is determined, the fitness threshold and the maximum number of iterations are pre-set, and the fitness value and ,when When the corresponding fitness value is higher than the fitness threshold, The threshold represented by is taken as the optimal threshold; otherwise, if the maximum number of iterations is reached, the number of iterations is set to 0 and the initial position is reset; otherwise, the number of iterations is increased by 1 and the neighborhood and individual position are rebuilt and updated, and the iteration continues;

[0080] The threshold processing unit designs a threshold processing function and performs threshold processing on the detail coefficients of each scale and each direction based on the optimal threshold. The formula used is as follows:

[0081] ;

[0082] In the formula, and are the original detail coefficients in the fth direction of the cth scale and the detail coefficients after threshold processing, c and f are the scale index and direction index respectively. is the optimal threshold in the fth direction of the cth scale, b is the adjustment coefficient in the range of (0, 5), sgn(•) is the sign function; when hour, ;when hour, ;when hour, The image reconstruction unit performs wavelet reconstruction on the original approximate coefficients and the detail coefficients after optimal threshold processing to obtain the optimized nuclear magnetic resonance image.

[0083] By performing the above operations, in view of the problem that various noises are generated during the nuclear magnetic resonance imaging process in the existing esophageal cancer nuclear magnetic resonance diagnosis system, resulting in blurring, poor contrast and missing details in the acquired esophageal cancer nuclear magnetic resonance images, and reducing the diagnostic accuracy of esophageal cancer, this scheme performs multi-scale wavelet decomposition on the nuclear magnetic resonance images, which helps to more accurately identify noise and effective nuclear magnetic resonance image information; the initial position of the chaotic sine correlation mapping is designed so that the initial position setting method is random but has a certain regularity, avoiding blindness; two neighborhoods are constructed based on the historical optimal position, and the individual position is updated in segments according to the reference neighbors in the neighborhood and combined with the cosine similarity to find the optimal threshold, and the position update direction is determined by the reference neighbors and similarity, so as to accelerate the process of finding the optimal threshold; the detail coefficient is thresholded based on the optimal threshold, and then the optimized nuclear magnetic resonance image is obtained by wavelet reconstruction, which removes noise more accurately, and at the same time retains the effective details of the nuclear magnetic resonance image to the greatest extent, improves the quality of the nuclear magnetic resonance image, and further improves the accuracy and reliability of esophageal cancer diagnosis.

[0084] Example 4, see Figure 1 and Figure 3This embodiment is based on the above embodiment. In the module for constructing the esophageal cancer nuclear magnetic resonance diagnosis model, a preliminary feature extraction unit, a global feature extraction unit, a local feature extraction unit, a feature fusion unit and an output unit are provided. Specifically, the following contents are included:

[0085] The preliminary feature extraction unit inputs the optimized nuclear magnetic resonance image of size A×B into the pyramid visual transformer PVTv2-B2 for processing to obtain a multi-scale feature map, specifically a The first characteristic graph E 1 , size is The second characteristic graph E 2 , size is The third characteristic graph E 3 and size The fourth characteristic graph E 4 , the first feature map E 1 is a low-level feature, the second feature map E 2 , the third characteristic graph E 3 And the fourth characteristic graph E 4 is a high-level feature; where A and B are the image height and width respectively;

[0086] Extract global feature units, perform two different aggregations on high-level features, and calculate the global features of the MRI image based on the two aggregation results, including the following:

[0087] The first aggregation, for the third feature map E 3 And the fourth characteristic graph E 4 To perform aggregation, the formula used is as follows:

[0088] ;

[0089] In the formula, R 1 It is the high-level feature of the first aggregation. 2 (•) is a 2x nearest neighbor upsampling operation, is the element-wise product operator, Concat (•) is the concatenation operation, and LE (•) is a series of operations consisting of a 3×3 convolutional layer, a GN normalization layer, and a SiLU activation function;

[0090] The second aggregation, for the second feature map E 2 , the third characteristic graph E 3 And the fourth characteristic graph E 4 To perform aggregation, the formula used is as follows:

[0091] ;

[0092] In the formula, R 2 It is the advanced feature of the second aggregation, Up 4(•) is a 4x nearest neighbor upsampling operation;

[0093] Based on R 1 and R 2 Get the global feature map, the formula used is as follows:

[0094] ;

[0095] In the formula, Q 1 is the global feature map;

[0096] Extract local feature units from the first feature map E 1 The local features of the nuclear magnetic resonance image are extracted from the image. The formula used is as follows:

[0097] ;

[0098] In the formula, Q 2 is the local feature map, C 1×1 (•) is a 1×1 convolution operation, SiLU (•) is a SiLU activation function, and GN (•) is a normalization operation;

[0099] The feature fusion unit selectively fuses global features and local features to obtain the final features of the MRI image, including the following:

[0100] Selective fusion, the formula used is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] Where σ(•) is the Sigmoid activation function, and minus D 2 and D 1 The reverse operation, is the element-wise addition operator, D 1 and D 2 They are the processed global feature map and local feature map, and Selective fusion features Figure 1 and selective fusion features Figure 2 ;

[0106] right and Connect to get the connection characteristics , a series of operations consisting of two 1×1 convolution layers, GN normalization layers, and SiLU activation functions are performed on the connection features to obtain the final features of the nuclear magnetic resonance image. The formula used is as follows:

[0107] ;

[0108] In the formula, , G is the final feature;

[0109] The output unit processes the final feature G through the fully connected layer and the Softmax function to obtain the probability vector Y, and selects the data label with the highest probability as the final output diagnosis result. The formula used is as follows:

[0110] ;

[0111] Where FC (•) is the fully connected layer and Softmax (•) is the normalized exponential function.

[0112] By performing the above operations, in view of the problem that the existing esophageal cancer MRI diagnosis system has esophageal cancer MRI images containing a large amount of detail information and semantic information, a single feature is difficult to fully describe the lesion information, and at the same time, esophageal cancer diagnosis needs to consider local lesions and the overall condition, lacks the comprehensive use of global and local information, and cannot accurately diagnose esophageal cancer, this scheme performs preliminary feature extraction on the optimized MRI images to obtain multi-scale feature maps, obtains esophageal cancer features at different levels and scales, and contains richer details and semantic information; performs two different aggregations on high-level features to extract global features, and captures global feature patterns and structural information more comprehensively through different aggregation methods, which helps to grasp the esophageal cancer condition as a whole; extracts local features from the first feature map to highlight local subtle features and abnormalities; selectively fuses global features and local features to obtain two selectively fused feature maps, connects and processes to obtain the final features, and outputs the diagnosis results, effectively combining the overall grasp of the global situation with the local fine information, so that the diagnosis results can reflect the overall condition of esophageal cancer and pay attention to specific local details, thereby improving the comprehensiveness and accuracy of esophageal cancer MRI diagnosis.

[0113] Example 5, see Figure 1 This embodiment is based on the above embodiment. In the diagnosis enhancement module, MRI images of patients to be diagnosed with esophageal cancer are collected. After the MRI images are optimized, they are input into the esophageal cancer MRI diagnosis model for processing. Based on the diagnosis results output by the model, the current status of the patients to be diagnosed with esophageal cancer is assisted in understanding.

[0114] 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.

[0115] 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.

[0116] 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. An artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer, characterized by: It includes image acquisition module, MRI image optimization module, esophageal cancer MRI diagnosis model building module and diagnosis enhancement module; The image acquisition module acquires historical MRI image data of patients with esophageal cancer; The NMR image optimization module performs multi-scale wavelet decomposition on the NMR image, designs the initial position of the chaotic sine correlation mapping, constructs two neighborhoods based on the historical optimal position, updates the individual position in segments based on the reference neighbors in the neighborhood and in combination with the cosine similarity, finds the optimal threshold, performs threshold processing on the detail coefficient based on the optimal threshold, and then performs wavelet reconstruction to obtain the optimized NMR image; The module for constructing the esophageal cancer nuclear magnetic resonance diagnosis model performs preliminary feature extraction on the optimized nuclear magnetic resonance image to obtain a multi-scale feature map, performs two different aggregations on the high-level features to extract global features, extracts local features from the first feature map, selectively fuses the global features and the local features to obtain two selectively fused feature maps, connects and processes them to obtain the final features, and outputs the diagnosis result; The diagnosis enhancement module assists in understanding the current status of the patient to be diagnosed with esophageal cancer based on the diagnosis results output by the esophageal cancer nuclear magnetic resonance diagnosis model; The nuclear magnetic image optimization module is provided with a threshold processing unit, and a threshold processing function is designed to perform threshold processing on the detail coefficients of each scale and each direction based on the optimal threshold. The formula used is as follows: ； In the formula, and are the original detail coefficients in the fth direction of the cth scale and the detail coefficients after threshold processing, c and f are the scale index and direction index respectively. is the optimal threshold in the fth direction at the cth scale, b is the adjustment coefficient, sgn(•) and exp(•) are the sign function and exponential function, respectively.

2. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 1, characterized in that: The nuclear magnetic resonance image optimization module is provided with an image decomposition unit, an optimal threshold search unit, a threshold processing unit and an image reconstruction unit, and specifically includes the following contents: The image decomposition unit uses db8 as the wavelet basis function to perform multi-scale wavelet decomposition on the nuclear magnetic image to obtain approximate coefficients and detail coefficients of multiple scales. The approximate coefficient is the low-frequency information of the nuclear magnetic image, and the detail coefficient is the high-frequency information of the nuclear magnetic image in the horizontal, vertical and diagonal directions; The optimal threshold search unit performs threshold processing on the detail coefficients of each scale and each direction, searches for the threshold based on the swarm intelligence algorithm, and finds the optimal threshold for each scale and each direction, including the following: Initial position: establish a search space based on the threshold of each scale and each direction, and randomly initialize the first individual position in the search space , design chaotic sinusoidal correlation mapping to initialize the remaining 2-I individual positions, the individual position dimension is equal to the number of scales × the number of directions, use the individual position as the representative of the threshold, and use the peak signal-to-noise ratio of the nuclear magnetic resonance image after threshold processing as the fitness value of the corresponding individual position; the formula used is as follows: ； In the formula, and are the initial positions of the i+1th and ith individuals in the jth dimension, respectively, i is the individual index, j is the dimension index, r1 is the first random number, r2 is the second random number, sin(•) is the sine function; Build a neighborhood and calculate the location of each individual and The Euclidean distance based on the closest The k individual positions of construct the first neighborhood of the i-th individual at the t-th iteration , based on the closest The k individual positions of construct the second neighborhood at the tth iteration ; where t is the iteration index, is the historical optimal position of the ith individual at the tth iteration, is the historical optimal position of individual I at the tth iteration. The historical optimal position is the position of the individual with the highest fitness value since the iteration; Individual location updates; The optimal threshold is determined, the fitness threshold and the maximum number of iterations are pre-set, and the fitness value and ,when When the corresponding fitness value is higher than the fitness threshold, The threshold represented by is taken as the optimal threshold; otherwise, if the maximum number of iterations is reached, the number of iterations is set to 0 and the initial position is reset; otherwise, the number of iterations is increased by 1 and the neighborhood and individual position are rebuilt and updated, and the iteration continues; Threshold processing unit; The image reconstruction unit performs wavelet reconstruction on the original approximate coefficients and the detail coefficients after processing based on the optimal threshold value to obtain an optimized nuclear magnetic resonance image.

3. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 2, characterized in that: The individual position in the optimal threshold search unit is updated, specifically from and Randomly select the first reference neighbor and the second reference neighbor ,calculate and The cosine similarity between , based on the cosine similarity segmentation, the individual position is updated, and the formula used is as follows: ； In the formula, and are the positions of the i-th individual in the j-th dimension at the t+1-th and t-th iterations, respectively. and They are and At the position of the jth dimension, and They are and at the position of the jth dimension, r3 and r4 are the third and fourth random numbers respectively, W t is the inertia weight at the tth iteration, and are the first acceleration coefficient and the second acceleration coefficient at the tth iteration, is the L2 norm.

4. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 1, characterized in that: The module for constructing the esophageal cancer nuclear magnetic resonance diagnosis model is provided with a preliminary feature extraction unit, a global feature extraction unit, a local feature extraction unit, a feature fusion unit and an output unit, and specifically includes the following contents: The preliminary feature extraction unit inputs the optimized nuclear magnetic resonance image of size A×B into the pyramid visual transformer PVTv2-B2 for processing to obtain a multi-scale feature map, specifically a The first feature map E1 of size The second feature map E2 of size The third feature map E3 and size is The fourth feature map E4, the first feature map E1 is a low-level feature, the second feature map E2, the third feature map E3 and the fourth feature map E4 are high-level features; wherein A and B are image height and width respectively; Extract global feature units, perform two different aggregations on high-level features, and calculate the global features of the MRI image based on the two aggregation results; Extract local feature units and extract local features of the nuclear magnetic resonance image from the first feature map E1. The formula used is as follows: ； Where Q2 is the local feature map, C 1×1 (•) is a 1×1 convolution operation, SiLU (•) is a SiLU activation function, GN (•) is a normalization operation, and LE (•) is a series of operations consisting of a 3×3 convolution layer, a GN normalization layer, and a SiLU activation function. is the element-wise product operator; Feature fusion unit; The output unit processes the final feature G through the fully connected layer and the Softmax function to obtain the probability vector Y, and selects the data label with the highest probability as the final output diagnosis result. The formula used is as follows: ； Where FC (•) is the fully connected layer and Softmax (•) is the normalized exponential function.

5. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 4, characterized in that: The extraction of the global feature unit specifically includes the following contents: In the first aggregation, the third feature map E3 and the fourth feature map E4 are aggregated, and the formula used is as follows: ； Where R1 is the high-level feature of the first aggregation, Up2 (•) is the 2x nearest neighbor upsampling operation, and Concat (•) is the concatenation operation; The second aggregation is to aggregate the second feature map E2, the third feature map E3 and the fourth feature map E4. The formula used is as follows: ； Where R2 is the high-level feature of the second aggregation, Up4 (•) is the 4-fold nearest neighbor upsampling operation; The global feature map is obtained based on R1 and R2, and the formula used is as follows: ； Where Q1 is the global feature map.

6. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 4, characterized in that: The feature fusion unit specifically includes the following contents: Selective fusion, the formula used is as follows: ； ； ； ； Where σ(•) is the Sigmoid activation function, and They are the reverse operations of subtracting D2 and D1, respectively. is an element-by-element addition operator, D1 and D2 are the processed global feature map and local feature map respectively, and They are selective fusion feature map 1 and selective fusion feature map 2 respectively; right and Connect to get the connection characteristics , a series of operations consisting of two 1×1 convolution layers, GN normalization layers, and SiLU activation functions are performed on the connection features to obtain the final features of the nuclear magnetic resonance image. The formula used is as follows: ； Where G is the final feature.

7. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 1, characterized in that: The image acquisition module collects historical esophageal cancer patient nuclear magnetic resonance image data; the historical esophageal cancer patient nuclear magnetic resonance image data includes nuclear magnetic resonance images and diagnosis results; the diagnosis results include normal, early, middle and late stages, and the diagnosis results are used as data labels.

8. The artificial intelligence-based enhanced nuclear magnetic resonance diagnosis system for esophageal cancer according to claim 1, characterized in that: The diagnosis enhancement module collects MRI images of patients to be diagnosed with esophageal cancer, optimizes the MRI images, and then inputs them into the esophageal cancer MRI diagnosis model for processing. Based on the diagnosis results output by the model, it assists in understanding the current status of the patients to be diagnosed with esophageal cancer.

Images