An artificial intelligence-assisted medical diagnosis method

Abstract

The present invention relates to the field of medical diagnosis, and in particular to an artificial intelligence-assisted medical diagnosis method. The content includes: collecting medical images and preprocessing them, generating a set of feature units, and extracting multidimensional feature vectors; constructing a dynamic relationship matrix based on the multidimensional feature vectors, and optimizing the multidimensional feature vectors and the dynamic relationship matrix; weighting the optimized multidimensional feature vectors through an adaptive feature interaction weighting algorithm to obtain weighted eigenvalues; classifying the feature units based on the weighted eigenvalues ​​to obtain classification and discrimination results; integrating the classification and discrimination results to generate a diagnosis report to assist medical diagnosis. It solves the problem that in the process of feature extraction and modeling, the subtle features of the lesion area cannot be fully captured; there is often an overfitting problem when processing feature units, which affects the generalization ability of the diagnosis results; the dynamic relationship between feature units cannot be fully combined, which limits the classification accuracy of complex lesion areas.

Description

Technical Field

[0001] The present invention relates to the field of medical diagnosis, and in particular to an artificial intelligence-assisted medical diagnosis method. Background Art

[0002] Current AI-assisted medical diagnosis technology has made significant progress in the field of clinical diagnosis. By using technologies such as deep learning, image processing, and support vector machines (SVM), key features can be extracted from medical images such as CT, MRI, and X-rays, and automatic lesion area recognition can be achieved. Current technology combines conventional image preprocessing methods (such as denoising, edge detection, image segmentation, etc.) and widely uses models such as convolutional neural networks (CNN) and recurrent neural networks (RNN) to process medical image data. It has high recognition accuracy and can effectively detect lesion areas and provide diagnostic suggestions. The combination of AI diagnosis systems and doctors' experience has become an increasingly important auxiliary tool, which has improved diagnostic efficiency and accuracy to a certain extent.

[0003] Although the above technologies have shown great potential in medical imaging diagnosis, there are still some problems that limit the effectiveness of their practical applications: first, they lack flexibility when dealing with complex lesion areas and cannot fully capture the subtle features of the lesion area, resulting in low recognition accuracy; second, there is often an overfitting problem when processing feature units, which affects the generalization ability of the diagnosis results; in addition, the feature weighting processing method is relatively simple and cannot fully combine the dynamic relationship between feature units, limiting the classification accuracy of complex lesion areas. The existence of the above problems affects the accuracy and reliability of medical image-assisted diagnosis. Summary of the invention

[0004] The present invention provides an artificial intelligence-assisted medical diagnosis method to solve the problems of lack of flexibility in the process of feature extraction and modeling when facing complex lesion areas, inability to fully capture the subtle features of the lesion area, resulting in low recognition accuracy; overfitting problems often exist when processing feature units, affecting the generalization ability of the diagnosis results; the feature weighting processing method is relatively single and cannot fully combine the dynamic relationship between feature units, limiting the classification accuracy of complex lesion areas.

[0005] An artificial intelligence-assisted medical diagnosis method of the present invention specifically includes the following technical solutions:

[0006] An artificial intelligence-assisted medical diagnosis method comprises the following steps:

[0007] S1: Collect medical images, pre-process the medical images, generate a set of feature units; extract multidimensional feature vectors based on the feature units; construct a dynamic relationship matrix based on the multidimensional feature vectors; introduce a two-step optimization mechanism to optimize the multidimensional feature vectors and the dynamic relationship matrix, and obtain the optimized dynamic relationship matrix and the optimized multidimensional feature vectors;

[0008] S2: Through the adaptive feature interaction weighting algorithm, the optimized multi-dimensional feature vector is weighted to obtain the weighted feature value; based on the weighted feature value, the feature units are classified to obtain the classification and discrimination results; the classification and discrimination results are integrated to generate a diagnosis report to assist medical diagnosis.

[0009] Preferably, the S1 specifically includes:

[0010] Based on the multidimensional feature vector, the dynamic relationship strength between feature units is calculated and a dynamic relationship matrix is ​​constructed. Based on the multidimensional feature vector and the dynamic relationship matrix, a molecular feature graph is constructed to model the feature units in the medical image.

[0011] Preferably, the S1 specifically includes:

[0012] The two-step optimization mechanism is divided into two stages: preliminary optimization and regularization optimization.

[0013] Preferably, the S1 specifically includes:

[0014] In the preliminary optimization stage, the dynamic relationship matrix and multi-dimensional feature vector are preliminarily optimized by minimizing the preliminary optimization objective function; the specific formula of the preliminary optimization objective function is as follows:

[0015]

[0016] Among them, J1 is the preliminary optimization objective function; ∈ is the smoothing parameter used to prevent the denominator from being zero; R kl is the characteristic unit U k and U l The dynamic relationship strength between them; k and l are the characteristic unit indices; F k and F l They are the characteristic units U k and U l The multidimensional feature vector of ; N is the total number of feature units; M is the total number of dimensions of the multidimensional feature vector; m is the dimension index of the multidimensional feature vector; ω m is the weight coefficient on the mth dimension; F km and F lm They are the characteristic units U k and U l The eigenvalue of the multidimensional eigenvector in the mth dimension.

[0017] Preferably, the S1 specifically includes:

[0018] On the basis of preliminary optimization, regularization optimization is carried out, and graph structure regularization terms and Gaussian kernel functions are introduced to construct the final optimization objective function; by minimizing the final optimization objective function, the optimized dynamic relationship matrix and the optimized multi-dimensional feature vector are obtained.

[0019] Preferably, the S2 specifically includes:

[0020] In the process of implementing the adaptive feature interaction weighted algorithm, a nonlinear attenuation mechanism is introduced to perform weighted processing on the optimized multidimensional feature vector to obtain the weighted feature value. The specific formula is as follows:

[0021]

[0022] in, is the weighted eigenvalue of the kth feature unit; M is the total number of dimensions of the multidimensional feature vector; N is the total number of feature units; m is the dimension index of the multidimensional feature vector; W m is the weighting coefficient of the mth dimension; is the eigenvalue of the optimized multidimensional feature vector of the kth feature unit in the mth dimension; is the eigenvalue of the optimized multidimensional feature vector of the lth feature unit in the mth dimension; is the dynamic relationship strength between the kth feature unit and the lth feature unit in the optimized dynamic relationship matrix; μ is the adjustment factor; γ is the nonlinear attenuation coefficient.

[0023] Preferably, the S2 specifically includes:

[0024] Based on the weighted eigenvalues, a classification discriminant function is introduced to classify the feature units to obtain a classification discriminant result; based on the classification discriminant result, it is determined whether the feature unit belongs to the lesion area.

[0025] Preferably, the S2 specifically includes:

[0026] The arithmetic mean is calculated for all weighted characteristic values ​​in the lesion area, and the severity of the lesion area is distinguished by setting the threshold interval. When the arithmetic mean falls into the high threshold interval, the lesion area is classified as a severe lesion; when the arithmetic mean falls into the middle threshold interval, the lesion area is classified as a moderate lesion; when the arithmetic mean falls into the low threshold interval, the lesion area is classified as a mild lesion.

[0027] The beneficial effects of the technical solution of the present invention are:

[0028] 1. The present invention models the feature units in medical images by constructing a "molecular feature map", which can clearly reflect the distribution and similarity of the feature units in the feature space. Compared with traditional image processing methods that rely on fixed grids or simple geometric models, the molecular feature map adaptively captures the complex relationships between feature units through a dynamic relationship matrix, making the artificial intelligence-assisted medical diagnosis system more flexible and efficient in dealing with the complex morphology of the lesion area, and can better identify the subtle differences in the lesions, thereby improving the detection accuracy and robustness of the lesion area, especially when facing fuzzy boundaries or complex lesion morphology.

[0029] 2. The two-step optimization mechanism ensures the reasonable distribution of feature units in the feature space through two stages: preliminary optimization and regularization optimization. The preliminary optimization stage enhances the correlation between feature units, allowing similar lesion areas to be more closely clustered, thereby improving the accuracy of identifying lesion areas; the regularization optimization stage avoids overfitting problems and makes the distribution of feature units have better generalization ability; the two-step optimization mechanism makes the diagnosis results more accurate and less susceptible to noise or data distribution bias, thereby improving the effectiveness of medical image-assisted diagnosis.

[0030] 3. The adaptive feature interaction weighted algorithm combines the weighted processing of multi-dimensional feature vectors with the feature adjustment of dynamic relationship matrices, which enhances the ability to capture the similarities and differences of feature units. It is suitable for complex or ambiguous lesion area recognition scenarios, and can dynamically adjust classification decisions according to the different characteristics of the lesion area, thereby improving the precision and accuracy of the recognition results and enhancing the accuracy of clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of an artificial intelligence-assisted medical diagnosis method described in the present invention. DETAILED DESCRIPTION

[0032] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in 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.

[0033] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0034] The following is a detailed description of a specific scheme of an artificial intelligence-assisted medical diagnosis method provided by the present invention in conjunction with the accompanying drawings.

[0035] Refer to the attached Figure 1 , which shows an artificial intelligence-assisted medical diagnosis method provided by an embodiment of the present invention, the method comprising the following steps:

[0036] S1: Collect medical images, pre-process the medical images, generate a set of feature units; extract multidimensional feature vectors based on the feature units; construct a dynamic relationship matrix based on the multidimensional feature vectors; introduce a two-step optimization mechanism to optimize the multidimensional feature vectors and the dynamic relationship matrix, and obtain the optimized dynamic relationship matrix and the optimized multidimensional feature vectors;

[0037] Collect medical images (such as CT, MRI or X-ray images) and preprocess the medical images, including standardization, edge detection and image segmentation. Specifically, use a Gaussian filter to denoise the medical image, and then perform grayscale standardization to obtain a standardized image; use the Canny edge detection algorithm to extract edge information from the standardized image to generate an edge image; use the region growing algorithm to segment the edge image to identify the region of interest (ROI) with the lesion probability area; then use the k-means clustering algorithm to divide the segmented region of interest ROI with the lesion probability area into feature units to generate a feature unit set Wherein N is the total number of feature units, which is set according to the specific implementation scenario, and each feature unit represents a lesion area or a part of a lesion area. The above preprocessing is a prior art and will not be described in detail here.

[0038] Furthermore, based on the feature unit U k , k ranges from 1 to N, and a multidimensional feature vector containing the spatial coordinates of the feature unit is extracted. In addition to the spatial coordinates of the feature unit, the type and quantity of the multidimensional features can be set according to different implementation scenarios. For example, according to the type of medical image, the nature of the lesion area, and the requirements of the diagnostic task, features of different dimensions such as spatial coordinates, grayscale, gradient, texture, etc. can be selected, and corresponding feature extraction methods can be used, such as using an edge detection algorithm to extract boundary features, using a gradient operator to calculate grayscale changes, using a local binary pattern (LBP) to extract texture features, or obtaining frequency domain information through Fourier transform, etc. The feature extraction method is a well-known method for those skilled in the art and will not be described in detail here.

[0039] Next, a molecular feature map is constructed based on the multidimensional feature vector to reflect the distribution of feature units in the feature space.

[0040] The molecular feature graph is a graph structure composed of a dynamic relationship matrix and a multidimensional feature vector. In the molecular feature graph, each feature unit is regarded as a molecule, and the multidimensional feature vector is the information of the molecule. Based on the multidimensional feature vector, the dynamic relationship strength between the feature units is calculated. The dynamic relationship strength is used to describe the dynamic relationship strength between molecules. The higher the similarity between molecules, the greater the dynamic relationship strength. Based on the dynamic relationship strength, a dynamic relationship matrix is ​​constructed.

[0041] The calculation formula for dynamic relationship strength is as follows:

[0042]

[0043] Among them, R kl is the characteristic unit U k and U l The dynamic relationship strength between them; k and l are the characteristic unit indices; F k and F l They are the characteristic units U k and U l is a multidimensional feature vector of ; α is the distance decay coefficient, which is used to control the rate at which the similarity between feature units decays as the distance increases; ||·|| is the operation of calculating the Euclidean distance; σ 2 is the normalized distance parameter, used for standardized distance calculation; distance attenuation coefficient α and normalized distance parameter σ 2 Obtained through experimental tuning.

[0044] A two-step optimization mechanism is applied to optimize the distribution of feature units in the feature space. The two-step optimization mechanism optimizes the dynamic relationship matrix and multi-dimensional feature vectors so that similar feature units can be more closely clustered in the feature space, thereby improving the accuracy of image recognition. The two-step optimization mechanism is divided into two stages: preliminary optimization and regularization optimization.

[0045] In the preliminary optimization stage, based on the dynamic relationship strength R between feature units kl and the multidimensional feature vector F k , the dynamic relationship matrix and multi-dimensional feature vector are preliminarily optimized by minimizing the preliminary optimization objective function. The preliminary optimization objective function introduces the weighted nonlinear distance calculation of the multi-dimensional feature vector, considers the weighted contribution of the feature vector in different dimensions, and enhances the nonlinear expression of feature differences, so as to capture complex feature relationships.

[0046] The specific formula of the preliminary optimization objective function is as follows:

[0047]

[0048] Among them, J1 is the preliminary optimization objective function; ∈ is a smoothing parameter, which is used to prevent the denominator from being zero and is set according to the specific implementation scenario; M is the total number of dimensions of the multidimensional feature vector and is set according to the specific implementation scenario; m is the dimension index of the multidimensional feature vector; ω m is the weight coefficient on the mth dimension, which is set according to the specific implementation scenario using expert experience; F km and F lm They are the characteristic units U k and U l The eigenvalue of the multidimensional eigenvector in the mth dimension;

[0049] On the basis of preliminary optimization, regularization optimization is performed; in the regularization optimization stage, a graph structure regularization term is added to construct the final optimization objective function. By minimizing the final optimization objective function, the feature units can avoid overfitting while maintaining similarity, ensuring that the distribution of feature units has good generalization ability. The final optimization objective function introduces graph structure regularization terms and Gaussian kernel functions to consider the local similarity between feature units and the rationality of global distribution; combined with nonlinear fractional expressions, the nonlinear difference expression between feature units is enhanced to ensure the smoothness and stability of the optimization process;

[0050] The specific formula of the final optimization objective function is as follows:

[0051]

[0052] Among them, h is the final optimization objective function; λ is the regularization parameter, which is used to control the influence of the graph structure regularization term on the final optimization objective function. It is obtained through experimental tuning to ensure that the final optimization avoids overfitting while maintaining the similarity of feature units. is the graph structure regularization term, where is a Gaussian kernel function, representing the feature unit U k and U l similarities between;

[0053] Through the two-step optimization mechanism, the optimized dynamic relationship matrix and the optimized multi-dimensional feature vector are obtained, which improves the correlation and distribution accuracy between feature units, accurately reflects the characteristics of the lesion area, and improves the diagnostic accuracy and robustness.

[0054] S2: Through the adaptive feature interaction weighting algorithm, the optimized multi-dimensional feature vector is weighted to obtain the weighted feature value; based on the weighted feature value, the feature units are classified to obtain the classification and discrimination results; the classification and discrimination results are integrated to generate a diagnosis report to assist medical diagnosis.

[0055] The optimized multi-dimensional feature vector is weighted using an adaptive feature interaction weighting algorithm to obtain a weighted feature value.

[0056] The weighted characteristic value is the comprehensive characteristic information of the characteristic unit, which is expressed in the form of a scalar and is used for subsequent calculation of classification and discrimination results to determine whether the characteristic unit belongs to a lesion area.

[0057] The adaptive feature interaction weighted algorithm is a weighted algorithm based on the optimized multi-dimensional feature vector and the optimized dynamic relationship matrix. It introduces a nonlinear attenuation mechanism, performs dynamic adjustments according to the eigenvalue differences between feature units, and outputs weighted eigenvalues, thereby enhancing the accurate capture of the similarities and differences of feature units.

[0058] The specific formula of the adaptive feature interaction weighted algorithm is as follows:

[0059]

[0060] in, is the weighted eigenvalue of the kth feature unit, expressed in scalar form; W m is the weighting coefficient of the mth dimension, reflecting the importance of each dimension in the classification, and is set according to the specific implementation scenario using expert experience; is the eigenvalue of the optimized multidimensional feature vector of the kth feature unit in the mth dimension; is the eigenvalue of the optimized multidimensional feature vector of the lth feature unit in the mth dimension; is the dynamic relationship strength between the kth feature unit and the lth feature unit in the optimized dynamic relationship matrix; μ is an adjustment factor used to adjust the interaction between feature units; γ is a nonlinear attenuation coefficient used to adjust the influence of the difference between the optimized multidimensional feature vectors on the weighted eigenvalue; μ and γ are obtained through experimental tuning;

[0061] Based on the weighted eigenvalues, the classification discriminant function is used to calculate the classification discriminant results, classify the feature units, and determine whether the feature units belong to the lesion area. The classification discriminant function introduces a nonlinear activation function based on the existing support vector machine (SVM) technology to perform classification and discrimination more accurately. The classification discriminant function is as follows:

[0062]

[0063] Among them, D(U k ) is the classification result. When the value is +1, it means the feature unit U k Belongs to the lesion area. When the value is -1, it means the characteristic unit U k belongs to the non-lesion area; sign is the sign function; θ m is the weight coefficient of the mth dimension; tanh is the hyperbolic tangent activation function; V mis the support vector of the mth dimension; b m is the bias term of the mth dimension, used to adjust the classification decision boundary; θ m 、V m 、b m are the parameters obtained by training the support vector machine.

[0064] After the classification and discrimination is completed, the classification and discrimination results are integrated to generate a diagnostic report including the overall distribution and severity of the lesion area.

[0065] The overall distribution of the lesion area, including the specific location and size of the lesion, is obtained based on the spatial coordinates of the characteristic units identified as the lesion area.

[0066] The severity of the lesion area is obtained by calculating the arithmetic mean of all weighted characteristic values ​​in the lesion area and then classifying them according to the high, medium and low threshold intervals set by the expert experience method: specifically, if the arithmetic mean falls into the high threshold interval, the lesion area is classified as a severe lesion; if the arithmetic mean falls into the medium threshold interval, the lesion area is classified as a moderate lesion; if the arithmetic mean falls into the low threshold interval, the lesion area is classified as a mild lesion. By setting the threshold interval, the severity of the lesion area can be distinguished, helping doctors make a comprehensive judgment on the severity of the lesion.

[0067] Finally, the diagnosis report is submitted to the doctor to assist in medical diagnosis, so that the doctor can confirm the diagnosis results based on clinical information and support subsequent treatment decisions.

[0068] In summary, an artificial intelligence-assisted medical diagnosis method is completed.

[0069] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0070] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. An artificial intelligence-assisted medical diagnosis method, characterized in that: The following steps are involved: S1: collect medical images, preprocess the medical images, and generate a set of feature units; Based on the feature unit, extract the multi-dimensional feature vector; Based on the multidimensional feature vector, a dynamic relationship matrix is ​​constructed; a double-step optimization mechanism is introduced to optimize the multidimensional feature vector and the dynamic relationship matrix to obtain the optimized dynamic relationship matrix and the optimized multidimensional feature vector; S2: The optimized multi-dimensional feature vector is weighted by an adaptive feature interaction weighting algorithm to obtain a weighted feature value; Based on the weighted eigenvalues, the feature units are classified to obtain the classification and discrimination results; Integrate classification and discrimination results, generate diagnostic reports, and assist in medical diagnosis.

2. The artificial intelligence-assisted medical diagnosis method according to claim 1, characterized in that: The S1 specifically includes: Based on the multidimensional feature vector, the dynamic relationship strength between feature units is calculated and a dynamic relationship matrix is ​​constructed. Based on the multidimensional feature vector and the dynamic relationship matrix, a molecular feature graph is constructed to model the feature units in the medical image.

3. The artificial intelligence-assisted medical diagnosis method according to claim 2, characterized in that: The S1 specifically includes: The two-step optimization mechanism is divided into two stages: preliminary optimization and regularization optimization.

4. The artificial intelligence-assisted medical diagnosis method according to claim 3, characterized in that: The S1 specifically includes: In the preliminary optimization stage, the dynamic relationship matrix and multi-dimensional feature vector are preliminarily optimized by minimizing the preliminary optimization objective function; the specific formula of the preliminary optimization objective function is as follows: Among them, J1 is the preliminary optimization objective function; ∈ is the smoothing parameter used to prevent the denominator from being zero; R kl is the characteristic unit U k and U l The dynamic relationship strength between them; k and l are the characteristic unit indices; F k and F l They are the characteristic units U k and U l The multidimensional feature vector of ; N is the total number of feature units; M is the total number of dimensions of the multidimensional feature vector; m is the dimension index of the multidimensional feature vector; ω m is the weight coefficient on the mth dimension; F km and F lm They are the characteristic units U k and U l The eigenvalue of the multidimensional eigenvector in the mth dimension.

5. The artificial intelligence-assisted medical diagnosis method according to claim 4, characterized in that: The S1 specifically includes: On the basis of preliminary optimization, regularization optimization is carried out, and graph structure regularization terms and Gaussian kernel functions are introduced to construct the final optimization objective function; by minimizing the final optimization objective function, the optimized dynamic relationship matrix and the optimized multi-dimensional feature vector are obtained.

6. The artificial intelligence-assisted medical diagnosis method according to claim 1, characterized in that: The S2 specifically includes: In the process of implementing the adaptive feature interaction weighted algorithm, a nonlinear attenuation mechanism is introduced to perform weighted processing on the optimized multidimensional feature vector to obtain the weighted feature value. The specific formula is as follows: in, is the weighted eigenvalue of the kth feature unit; M is the total number of dimensions of the multidimensional feature vector; N is the total number of feature units; m is the dimension index of the multidimensional feature vector; W m is the weighting coefficient of the mth dimension; is the eigenvalue of the optimized multidimensional feature vector of the kth feature unit in the mth dimension; is the eigenvalue of the optimized multidimensional feature vector of the lth feature unit in the mth dimension; is the dynamic relationship strength between the kth feature unit and the lth feature unit in the optimized dynamic relationship matrix; μ is the adjustment factor; γ is the nonlinear attenuation coefficient.

7. The artificial intelligence-assisted medical diagnosis method according to claim 6, characterized in that: The S2 specifically includes: Based on the weighted eigenvalues, a classification discriminant function is introduced to classify the feature units to obtain a classification discriminant result; based on the classification discriminant result, it is determined whether the feature unit belongs to the lesion area.

8. The artificial intelligence-assisted medical diagnosis method according to claim 7, characterized in that: The S2 specifically includes: The arithmetic mean is calculated for all weighted characteristic values ​​in the lesion area, and the severity of the lesion area is distinguished by setting the threshold interval. When the arithmetic mean falls into the high threshold interval, the lesion area is classified as a severe lesion; when the arithmetic mean falls into the middle threshold interval, the lesion area is classified as a moderate lesion; when the arithmetic mean falls into the low threshold interval, the lesion area is classified as a mild lesion.

Images