A medical manifold vector set classification method for thyroid nodules in an ultrasound image

Abstract

The application provides a medical manifold vector set classification method for thyroid nodules in an ultrasound image, comprising: preparing a thyroid nodule ultrasound image data set; designing and training a manifold mapping network to obtain high-dimensional vector coordinates of each pixel point of the thyroid nodule image; performing dimension reduction on the high-dimensional vector space based on Fourier transform; within a single nodule, based on spatial position information and high-dimensional vector similarity, performing pixel or superpixel merging from bottom to top to generate a manifold vector category distribution map corresponding to the nodule; based on the category distribution maps of all nodules and their manifold spaces, measuring the distance between different vector space manifolds in a global range, and merging vector sets with similar distances, to finally obtain the classification results of each vector set; and performing manifold vector set category distribution statistics for benign and malignant nodules respectively. The method has the advantages of less data labeling, rapidness, high interpretability, intuitive results, convenience for medical analysis and diagnosis, and the like.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a medical manifold vector set classification method for thyroid nodules in ultrasound images. Background Technology

[0002] Thyroid nodules are very common in clinical practice, and their incidence has been steadily increasing over the past two decades. This is partly due to advancements in medical testing technology, which have improved detection rates. Various medical imaging methods can now detect some asymptomatic small thyroid nodules. Among these, ultrasound examination has become the most commonly used method for thyroid nodule examination due to its advantages such as safety (no radiation), ease of operation, and low cost. Ultrasound examination diagnoses thyroid nodules by measuring important signs such as size, borders, shape, echogenicity, and blood flow information. In two-dimensional ultrasound, the more comprehensive the analysis of ultrasound signs, the higher the accuracy; however, it is also more time-consuming and labor-intensive. Furthermore, ultrasound examination has been limited in terms of diagnostic accuracy, standardization, and uniformity due to its reliance on the subjective judgment of the operating physician.

[0003] Existing methods for automatic analysis of thyroid nodule ultrasound image features mostly employ supervised learning to train multi-task classification networks, such as Chinese patents with publication numbers CN111341443A and CN112270667A. These methods require annotation of nodule echogenicity, composition, boundaries, and punctate strong echoes, making dataset creation extremely costly in terms of manpower and time. Other methods, such as the cystic component analysis method proposed in Chinese patent CN112102311A, only use pixel value information, limiting the accuracy of cystic and solid region segmentation and making it difficult to adapt to nodules with complex structures. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a medical manifold vector set classification method for thyroid nodules.

[0005] According to one aspect of the present invention, a medical manifold vector set classification method for thyroid nodules in ultrasound images is provided, comprising:

[0006] To obtain a dataset of ultrasound images of thyroid nodules;

[0007] A manifold mapping network is used to obtain the high-dimensional vector coordinates of each pixel in the ultrasound image of the thyroid nodule, thus obtaining a high-dimensional vector space.

[0008] The high-dimensional vector space is reduced in dimensionality based on Fourier transform;

[0009] Within a single nodule, based on vector similarity, hierarchical clustering is used to merge pixels or superpixels from bottom to top in the dimensionality-reduced high-dimensional vector space to form a vector set, generating a manifold vector category distribution map corresponding to the nodule;

[0010] Based on the manifold vector category distribution map of all nodules, the distance between different vector space vector sets is measured globally, and vector sets with close distances are merged to finally obtain the classification results of each vector set;

[0011] Based on the classification results of each vector set, statistical analysis of the distribution of benign and malignant nodules was performed.

[0012] Preferably, the structure of the manifold mapping network includes an encoder part and a decoder part, and is capable of accepting ultrasound images of thyroid nodules of any size as input;

[0013] The encoder part includes multiple convolutional modules, each of which performs one or more of the following operations: convolution, batch normalization, activation function, and pooling; the strides of each convolutional module are all 1; the multiple convolutional modules are dilated convolutions of different scales, which can obtain different receptive fields;

[0014] The decoder part is symmetrical to the encoder part and includes multiple deconvolution modules. Each deconvolution module performs one or more of the following operations: convolution, batch normalization, activation function, and pooling. The strides of each deconvolution module are 1.

[0015] The output of the encoder is a high-dimensional vector space containing multi-dimensional nodal feature maps with receptive fields of different sizes, and the size of the feature map is the same as the input size of the manifold mapping network.

[0016] Preferably, the dimensionality reduction of the high-dimensional vector space based on Fourier transform includes:

[0017] Statistically determine the proportion of non-zero values ​​in each dimension of a high-dimensional vector space, and eliminate coordinate dimensions whose proportion of non-zero values ​​is lower than a set threshold;

[0018] Obtain the two-dimensional Fourier transform spectrum of the feature map corresponding to each vector dimension, and calculate the score for each dimension;

[0019] Calculate the Pearson correlation coefficient between the two feature maps based on the spectrum;

[0020] For a nodule image, the optimal n-dimensional vector space of the nodule is obtained by combining the scores of all dimensions and the Pearson correlation coefficient.

[0021] Preferably, the step of obtaining the two-dimensional Fourier transform spectrum of the feature map corresponding to the vector dimension and calculating the score of each dimension includes:

[0022] The scoring method for each dimension is as follows:

[0023] Score = Score lf +αScore hf

[0024]

[0025]

[0026] Among them, Score lf Score represents the score for the low-frequency region. hf Let α be the high-frequency region fraction, FL be the low-frequency region matrix extracted from the spectrum, FH be the high-frequency region matrix extracted from the spectrum, and n1 and n2 be the dimensions of the FL or FH matrix.

[0027] Preferably, the step of deriving the optimal n-dimensional vector space for the nodule image by combining the scores of all dimensions and the Pearson correlation coefficient includes:

[0028] For each nodule image, the dimension with the smallest dimensional score is selected as the starting dimension;

[0029] Using the absolute value of the Pearson correlation coefficient as the correlation coefficient between vector dimensions, the correlation coefficients (Score) of the remaining dimensions in the vector space with the initial dimension are calculated. r Select Score+βScore r The smallest dimension, where β is a constant;

[0030] Calculate the average of the correlation coefficients between the remaining dimensions and the selected dimensions in the vector space. choose The smallest dimension is selected; repeat this step until the vector space has selected the preset number of dimensions.

[0031] Preferably, within a single nodule, based on vector similarity, hierarchical clustering is used from bottom to top to merge pixels or superpixels in the dimensionality-reduced high-dimensional vector space to generate a manifold vector category distribution map corresponding to the nodule, including:

[0032] Euclidean distance is used as the vector similarity;

[0033] Number all pixels within a single nodule as the initial category, and obtain point sets of multiple categories;

[0034] For each category's point set, obtain the neighboring category numbers;

[0035] For each category's point set, calculate the distance between that category and all its neighboring categories, find and record the category number with the smallest distance;

[0036] If two categories are the categories with the smallest distance to each other, then the two categories are considered to be matched successfully, and the matched category pairs are merged and renumbered;

[0037] Repeat the above operation until the termination condition is met: the number of adjacent categories is less than a certain specified threshold, or the total number of categories in the nodule is less than a certain specified threshold, or there are no successfully matched categories.

[0038] Once the above termination conditions are met, the AGNES method is used to perform hierarchical clustering to reduce the number of categories to the target number.

[0039] Preferably, the manifold vector category distribution map based on all nodes measures the distance between different vector space vector sets globally, merges vector sets with similar distances, and finally obtains the classification results for each vector set, including:

[0040] A set number of samples are randomly selected from all vector sets for clustering.

[0041] Design methods for measuring vector set distance and class distance in different vector spaces;

[0042] A bottom-up hierarchical clustering method is used to merge categories until the global number of categories is reduced to a specified threshold;

[0043] Classify vector sets outside the sample based on the distance between vector sets.

[0044] Preferably, the method for measuring the distance between vector sets in different vector spaces includes:

[0045] For vector set ps i Calculate the distance between it and each of the other vector sets in the sample, and record the minimum and maximum distances:

[0046]

[0047] Among them, ps i and ps j For two different vector sets, fl i and fl j They are vector sets ps i and ps j List of dimensions in vector space, Coordinate(ps) j ,fl i ) represents the vector set ps j In fl i Coordinate matrix in dimension Represents the vector set ps j In fl i The coordinates of the center point in the dimension Represents the vector set ps i and ps j In vector set ps i The distance in the corresponding vector space dimension; mind i and maxd i They represent the dimensions fl in the vector space, respectively. i Below, vector set ps i Minimum and maximum distances between the vector set and the rest of the vector set.

[0048] Preferably, the method for measuring class distance in different vector spaces includes:

[0049] For category This includes one or more vector sets belonging to the same or different nodules, whose values ​​are calculated relative to the category. Distance:

[0050]

[0051] Preferably, classifying vector sets outside the sample based on vector set distance includes:

[0052] In the corresponding vector space dimension of the vector set outside the sample, calculate the distance between it and all vector sets in the sample, and select the global category to which the vector set with the smallest distance belongs as its classification result.

[0053] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0054] The medical manifold vector set classification method for thyroid nodules in ultrasound images described in this invention provides advantages such as masked annotation of thyroid nodules, low data output, high speed, and intuitive results. This invention employs a clustering method, which is an unsupervised learning approach. The distance calculation method between vectors is manually designed to ensure the accuracy and reliability of the classification. Compared to existing supervised neural network classification methods, this method avoids the limitations of black-box methods and provides higher interpretability. Attached Figure Description

[0055] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0056] Figure 1 This is a flowchart of a medical manifold vector set classification method for thyroid nodules according to a preferred embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of a manifold mapping network structure in a preferred embodiment of the present invention;

[0058] Figure 3 This is an example diagram of the distribution of manifold vector categories within a single nodule in a preferred embodiment of the present invention;

[0059] Figure 4 This is an example diagram of the global vector set clustering result in a preferred embodiment of the present invention. Detailed Implementation

[0060] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0061] Figure 1 This is a flowchart of a medical manifold vector set classification method for thyroid nodules in a preferred embodiment of the present invention, specifically including the following steps:

[0062] S1, a dataset of ultrasound images of thyroid nodules;

[0063] S2, using a manifold mapping network, the high-dimensional vector coordinates of each pixel in the ultrasound image of the thyroid nodule are obtained to obtain a high-dimensional vector space;

[0064] S3, Dimensionality reduction of the high-dimensional vector space based on Fourier transform;

[0065] S4. Within a single node, based on vector similarity, hierarchical clustering is used to merge pixels or superpixels (a superpixel is a pixel block composed of several pixels, that is, the image is transformed from pixel level to region level) from bottom to top in the dimensionality-reduced high-dimensional vector space to generate a vector set, and obtain the manifold vector category distribution map corresponding to the node. Each manifold vector category includes at least one vector set.

[0066] S5, based on the distribution map of manifold vector categories of all nodes and their manifold spaces, measures the distance between different vector space manifolds in the global scope, merges vector sets with close distances, and finally obtains the classification results of each vector set;

[0067] S6 performs statistical analysis of the distribution of manifold vector sets for benign and malignant nodules, respectively.

[0068] In a preferred embodiment of the present invention, the dataset of S1 includes the original thyroid ultrasound image, thyroid nodule mask, and annotations of the true values ​​of benign and malignant thyroid nodules.

[0069] In a preferred embodiment of the present invention, the manifold mapping network structure of S2 is as follows: Figure 2As shown, the network includes an encoder and a decoder, using dilated convolutions of different scales to acquire features within different receptive fields. The specific structure of the network is as follows:

[0070] The first layer consists of a convolutional module with a window size of 5×5, 16 output channels, a stride of 1, and a dilation rate of 1.

[0071] The second layer connects the output of the first layer to convolutional modules with a window size of 5×5, 32 output channels, a stride of 1, and dilation rates of 1 and 2, respectively.

[0072] The third layer connects the output of the second-layer convolutional module with a dilation rate of 1 to a convolutional module with a window size of 5×5, 64 output channels, a stride of 1, and a dilation rate of 1. The output of the second-layer convolutional module with a dilation rate of 2 connects to convolutional modules with a window size of 5×5, 64 output channels, a stride of 1, and dilation rates of 2 and 3 respectively.

[0073] The three sets of features obtained from the encoding are the outputs of branches with three different receptive fields, namely group1, group2, and group3. The size of each set of features is the same as the original image size, and the number of features is 64.

[0074] All of the above convolutional modules use batch normalization layers and ReLU activation functions, and 2×2 and 4×4 pooling layers are connected after the dilated convolutional modules with strides of 2 and 3, respectively.

[0075] In the fourth layer, feature group 1 is processed by a deconvolution module with a window size of 5×5, 32 output channels, a stride of 1, and a dilation rate of 1; feature group 2 is processed by a deconvolution module with a window size of 5×5, 32 output channels, a stride of 1, and a dilation rate of 2; and feature group 3 is processed by a deconvolution module with a window size of 5×5, 32 output channels, a stride of 1, and a dilation rate of 3.

[0076] The fifth layer, the output of the convolutional module with a dilation rate of 1 in the fourth layer is connected to a deconvolutional module with a window size of 5×5, 16 output channels, a stride of 1, and a dilation rate of 1. The outputs of the convolutional module with a dilation rate of 2 and the output of the convolutional module with a dilation rate of 3 in the fourth layer are concatenated and then connected to a deconvolutional module with a window size of 5×5, 16 output channels, a stride of 1, and a dilation rate of 2.

[0077] The sixth layer concatenates the two outputs from the fifth layer and connects them to a deconvolution module with a window size of 5×5, 1 output channel, 1 stride, and 1 dilation rate, ultimately outputting the reconstructed image.

[0078] The aforementioned intermediate deconvolutional modules all use batch normalization layers and ReLU activation functions, while the output layer uses linear activation functions.

[0079] The core function of manifold mapping networks lies in achieving accurate mapping from low-dimensional to high-dimensional spaces. For single-channel grayscale images of thyroid nodules—i.e., one-dimensional data—this embodiment employs a manifold mapping network to effectively transform them into a high-dimensional space, ensuring that each pixel corresponds to an accurate high-dimensional vector. Given the limitations of traditional feature extraction methods in feature mining and noise reduction, this embodiment specifically introduces a manifold mapping network to extract richer and more accurate features, thereby optimizing subsequent data analysis and processing.

[0080] Ultrasound images inherently possess speckle noise. The aforementioned manifold mapping network maps a one-dimensional space to a high-dimensional space, which includes not only valid information but also a significant amount of noise. Some dimensions are heavily influenced by noise, hindering clustering, necessitating the selection of suitable dimensions. Therefore, in a preferred embodiment of the present invention, implementation S3 involves dimensionality reduction of the high-dimensional vector space based on Fourier transform, including the following steps:

[0081] Statistically determine the proportion of non-zero values ​​in each dimension and eliminate coordinate dimensions where the proportion of non-zero values ​​is less than 80%.

[0082] For each nodule image, calculate the two-dimensional Fourier transform spectrum of the feature map corresponding to all vector dimensions, and calculate the score for each dimension using the following formula:

[0083] Score = Score lf +α·Score hf

[0084]

[0085] Among them, Score lf Score represents the score for the low-frequency region. hf Let α be the high-frequency region fraction, FL be the low-frequency region matrix extracted from the spectrum, FH be the high-frequency region matrix extracted from the spectrum, and n1 and n2 be the dimensions of the FL or FH matrix.

[0086] For each nodule image, the dimension with the smallest score in each set of vector spaces is selected as the starting dimension;

[0087] The absolute value of the Pearson correlation coefficient of the feature spectrogram corresponding to the vector dimension is used as the correlation coefficient between vector dimensions. The correlation coefficients (Score) of the other dimensions in the vector space with the initial dimension are then calculated. r Select Score+βScore r The smallest dimension, where β is 5;

[0088] Similarly, calculate the average of the correlation coefficients between the remaining dimensions and the selected dimensions in the vector space. choose The minimum dimension, where β is 5, is selected. This step is repeated until the vector space has 5 selected dimensions.

[0089] Existing techniques typically extract features from one-dimensional images, making it difficult to eliminate the influence of noise. The above embodiment maps the one-dimensional space to a high-dimensional space, and then obtains a vector space suitable for clustering by evaluating each dimension, which can greatly reduce the impact of noise on clustering.

[0090] In a preferred embodiment of the present invention, step S4 involves performing pixel or superpixel merging from bottom to top within a single nodule based on spatial location information and high-dimensional vector similarity to generate a manifold vector category distribution map corresponding to that nodule, including the following steps:

[0091] Number all pixels within a single nodule as the initial category;

[0092] For each category's point set, obtain its four neighboring category numbers;

[0093] Calculate two point sets using the following formula. and Distance between:

[0094]

[0095] Among them, psd 12 This represents the distance between two point sets within the same node. This represents the Euclidean distance between two points within the same nodule.

[0096] For each category's vector set, calculate the distance between that category and all its neighboring categories, find and record the category number with the smallest distance;

[0097] To speed up clustering, the distance between the center points of two vector sets is used as a metric for the distance between them in the later stages of clustering.

[0098] If two categories are the categories with the smallest distance to each other, then the two categories are considered to be matched successfully, and the matched category pairs are merged and renumbered;

[0099] Repeat the above operation until the termination condition is met: the number of adjacent categories is less than 3 or there are no successfully matched categories;

[0100] Once the above termination condition is met, hierarchical clustering is performed using Euclidean distance as the distance metric and the AGNES method until the number of categories is reduced to 30, generating the manifold vector category result map for the nodule.

[0101] like Figure 3 The figure shows an example of the distribution of manifold vector categories within a single node. The same color represents the same category, and the number of manifold vector categories in each node is 30.

[0102] In a preferred embodiment of the present invention, step S5 involves measuring the distance between different vector space manifolds globally based on the nodule category distribution map and its manifold space, and merging vector sets with similar distances to obtain the classification results for each vector set. The implementation steps are as follows:

[0103] 1000 vector sets are randomly selected from all vector sets as samples;

[0104] All vector sets in the sample are assigned a category number, and the distance between each vector set and all other vector sets in the sample is calculated using the following formula. And record mind i and maxd i :

[0105]

[0106] Among them, ps i and ps j For two different vector sets, fl i and fl j They are vector sets ps i and ps j List of dimensions in vector space, Coordinate(ps) j ,fl i ) represents the vector set ps j In fl i Coordinate matrix in dimension Represents the vector set ps j In fl i The coordinates of the center point in the dimension Represents the vector set ps i and ps j In vector set ps i The distance in the corresponding vector space dimension. i and maxd i They represent the dimensions fl in the vector space, respectively. i Below, vector set ps i Minimum and maximum distances between the vector set and the rest of the vector set;

[0107] If two vector sets are the vector sets with the smallest distance to each other, then the two vector sets are considered to be matched successfully, and the matched vector sets are merged and renumbered.

[0108] Calculate the category using the following formula. With category Distance:

[0109]

[0110] After calculating the distance between each category, hierarchical clustering is performed using the AGNES method until the number of global categories is reduced to a specified threshold.

[0111] For vector sets outside the samples, in the spatial dimension of the vector set, calculate the distance between it and all vector sets in the samples, and select the global category to which the vector set with the smallest distance belongs as its classification result.

[0112] like Figure 4 This is an example diagram of the global vector set clustering results. The diagram shows two categories, and each category lists the corresponding vector set mask in three nodes.

[0113] On the experimental test set, the method achieved a classification accuracy of 96.8% for cystic / solid lesions, 86.6% for hypoechoic / isoechoic / hyperechoic lesions, and 95.8% for calcified / non-calcified lesions.

[0114] Compared with existing technologies, the medical manifold vector set classification method for thyroid nodules in ultrasound images provided by the above embodiments of the present invention obtains feature dimensions under different receptive fields to form a high-dimensional coordinate space through a manifold mapping network; then, an unsupervised method is used for dimensionality reduction; furthermore, vector set clustering within different nodules is achieved through hierarchical clustering within individual nodules and global hierarchical clustering; finally, the vector set category characteristics within benign and malignant nodules can be analyzed based on the clustering results. This method has advantages such as requiring less data annotation, speed, high interpretability, intuitive results, and ease of medical analysis and diagnosis.

[0115] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A medical manifold vector set classification method for thyroid nodules in ultrasound images, characterized in that, include: To obtain a dataset of ultrasound images of thyroid nodules; A manifold mapping network is used to obtain the high-dimensional vector coordinates of each pixel in the ultrasound image of the thyroid nodule, thus obtaining a high-dimensional vector space. The high-dimensional vector space is reduced in dimensionality based on Fourier transform; Within a single nodule, based on vector similarity, hierarchical clustering is used to merge pixels or superpixels from bottom to top in the dimensionality-reduced high-dimensional vector space to form a vector set, generating a manifold vector category distribution map corresponding to the nodule; Based on the manifold vector category distribution map of all nodules, the distance between different vector space vector sets is measured globally, and vector sets with close distances are merged to finally obtain the classification results of each vector set; Based on the classification results of each vector set, statistical analysis of the distribution of benign and malignant nodules was performed. Specifically, within a single nodule, based on vector similarity, hierarchical clustering is used from bottom to top to merge pixels or superpixels in the dimensionality-reduced high-dimensional vector space to generate a manifold vector category distribution map corresponding to that nodule, including: Euclidean distance is used as the vector similarity; Number all pixels within a single nodule as the initial category, and obtain point sets of multiple categories; For each category's point set, obtain the neighboring category numbers; For each category's point set, calculate the distance between that category and all its neighboring categories, find and record the category number with the smallest distance; If two categories are the categories with the smallest distance to each other, then the two categories are considered to be matched successfully, and the matched category pairs are merged and renumbered; Repeat the above operation until the termination condition is met: the number of adjacent categories is less than a certain specified threshold, or the total number of categories in the nodule is less than a certain specified threshold, or there are no successfully matched categories. Once the above termination conditions are met, the AGNES method is used to perform hierarchical clustering to reduce the number of categories to the target number, generating a manifold vector category result map for the nodule.

2. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 1, characterized in that, The structure of the manifold mapping network includes an encoder part and a decoder part, and it can accept ultrasound images of thyroid nodules of any size as input; The encoder part includes multiple convolutional modules, each of which performs one or more of the following operations: convolution, batch normalization, activation function, and pooling; the strides of each convolutional module are all 1; the multiple convolutional modules are dilated convolutions of different scales, which can obtain different receptive fields; The decoder part is symmetrical to the encoder part and includes multiple deconvolution modules. Each deconvolution module performs one or more of the following operations: convolution, batch normalization, activation function, and pooling. The strides of each deconvolution module are 1. The output of the encoder is a high-dimensional vector space containing multi-dimensional nodal feature maps with receptive fields of different sizes, and the size of the feature map is the same as the input size of the manifold mapping network.

3. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 1, characterized in that, The dimensionality reduction of the high-dimensional vector space based on Fourier transform includes: Statistically determine the proportion of non-zero values ​​in each dimension of a high-dimensional vector space, and eliminate coordinate dimensions whose proportion of non-zero values ​​is lower than a set threshold; Obtain the two-dimensional Fourier transform spectrum of the feature map corresponding to each vector dimension, and calculate the score for each dimension; Calculate the Pearson correlation coefficient between the two feature maps based on the spectrum; For a nodule image, the optimal n-dimensional vector space of the nodule is obtained by combining the scores of all dimensions and the Pearson correlation coefficient.

4. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 3, characterized in that, The process of obtaining the two-dimensional Fourier transform spectrum of the feature map corresponding to the vector dimension and calculating the score for each dimension includes: The scoring method for each dimension is as follows: ； ； ； in, The score represents the low-frequency region. For high-frequency region scores, are real constants. This is a matrix of low-frequency regions extracted from the spectrum. This is a matrix of high-frequency regions extracted from the spectrum. and for Matrix or The size of the matrix.

5. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 3, characterized in that, The process of deriving the optimal n-dimensional vector space for a nodule image by combining the scores of all dimensions and the Pearson correlation coefficient includes: For each nodule image, the dimension with the smallest dimensional score is selected as the starting dimension; Using the absolute value of the Pearson correlation coefficient as the correlation coefficient between vector dimensions, calculate the correlation coefficients between the remaining dimensions in the vector space and the initial dimension. ,choose The smallest dimension, of which It is a constant; Calculate the average of the correlation coefficients between the remaining dimensions and the selected dimensions in the vector space. ,choose The smallest dimension is selected. This step is repeated until the number of dimensions selected in the vector space reaches the preset number.

6. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 1, characterized in that, The manifold vector category distribution map based on all nodes measures the distance between different vector sets in the vector space globally, merges vector sets with similar distances, and finally obtains the classification results for each vector set, including: A set number of samples are randomly selected from all vector sets for clustering. Design methods for measuring vector set distance and class distance in different vector spaces; A bottom-up hierarchical clustering method is used to merge categories until the global number of categories is reduced to a specified threshold; Classify vector sets outside the sample based on the distance between vector sets.

7. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 6, characterized in that, Methods for measuring the distance between vector sets in different vector spaces include: For vector sets Calculate the distance between it and each of the other vector sets in the sample, and record the minimum and maximum distances: ； ； ； ； in, and For two different vector sets, and Vector sets and List of dimensions in vector space Represents a vector set exist Coordinate matrix in dimension Represents a vector set exist The coordinates of the center point in the dimension Represents a vector set and In vector set The distance in the corresponding vector space dimension; and These respectively represent the dimensions in the vector space. Below, vector set Minimum and maximum distances between the vector set and the rest of the vector set.

8. The medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 6, characterized in that, Methods for measuring class distance in different vector spaces include: For category This includes one or more vector sets belonging to the same or different nodules, whose values ​​are calculated relative to the category. Distance: 。 9. A medical manifold vector set classification method for thyroid nodules in ultrasound images according to claim 6, characterized in that, The classification of vector sets outside the sample based on vector set distance includes: In the corresponding vector space dimension of the vector set outside the sample, calculate the distance between it and all vector sets in the sample, and select the global category to which the vector set with the smallest distance belongs as its classification result.

Images