Thyroid lymphoma prediction method and device based on machine learning

Abstract

The present invention relates to a thyroid lymphoma prediction method and device based on machine learning, which belongs to the field of image processing technology and solves the problem of the lack of thyroid lymphoma prediction methods based on deep learning in the prior art. The specific steps include: annotating the thyroid ultrasound image with the lesion area segmented, including the lesion type label of lymphoma, to obtain the original data set; extracting the tumor imaging features of the original data set to obtain the initial features, and performing zero importance analysis and correlation screening in sequence to obtain the final screening features with high importance and low correlation; based on the final screening features and the annotated lesion type labels, using machine learning training to obtain a thyroid lymphoma prediction model; performing feature extraction on the thyroid ultrasound segmentation image to be tested with the suspected lesion area segmented, to obtain the corresponding final screening features; based on the corresponding final screening features, using the thyroid lymphoma prediction model to achieve accurate prediction of thyroid lymphoma.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a thyroid lymphoma prediction method and device based on machine learning. Background Art

[0002] Thyroid lymphoma is a malignant tumor that poses a significant threat to human health. Not only is it associated with high morbidity and mortality, but its treatment is also costly and complex, placing a heavy financial and psychological burden on patients and their families. Traditional treatments primarily include chemotherapy, radiotherapy, and surgical resection. However, due to the lack of obvious early symptoms, patients are often diagnosed in the late stages of the disease, by which time the tumor may have extensively invaded surrounding tissues or metastasized to distant sites. This results in poor treatment efficacy, a poor prognosis, and a low survival rate. Therefore, accurately predicting the risk of thyroid lymphoma at an early stage, enabling early detection, diagnosis, and treatment, is crucial for improving patient cure rates and outcomes. In the field of medical imaging, ultrasound examination is the preferred method for screening and diagnosing thyroid diseases, offering advantages such as being non-invasive, convenient, and highly reproducible. However, relying solely on morphological features in ultrasound images for the diagnosis of thyroid lymphoma has certain limitations, making it difficult to accurately distinguish between benign nodules and malignant lymphomas, which can easily lead to misdiagnosis or missed diagnosis. In recent years, with the rapid development of deep learning technology, its application in medical image analysis has become increasingly extensive, providing new ideas and methods for the early diagnosis and prediction of thyroid lymphoma.

[0003] In the clinical research and practice of thyroid lymphoma, precise segmentation and accurate diagnosis of lesions are crucial first steps. Researchers have made significant progress, particularly in using deep learning techniques to assist in the analysis of ultrasound images. This has enabled more precise identification and delineation of thyroid lymphoma, playing a significant role in improving diagnostic accuracy and efficiency. However, there is currently relatively little deep learning research on thyroid lymphoma, especially in the use of deep learning techniques for disease prediction, which is still in its infancy. There is still a lack of systematic research and effective solutions for using deep learning models to predict individual disease risk, achieve early screening, and identify high-risk populations. Summary of the Invention

[0004] In view of the above analysis, the embodiments of the present invention aim to provide a thyroid lymphoma prediction method and device based on machine learning to address the problem of the lack of thyroid lymphoma prediction based on deep learning in the prior art.

[0005] The purpose of the present invention is mainly achieved through the following technical solutions:

[0006] In one aspect, an embodiment of the present invention provides a method for predicting thyroid lymphoma based on machine learning, comprising the following steps:

[0007] The thyroid ultrasound images with segmented lesion areas are labeled with lesion type labels including lymphoma to obtain the original dataset;

[0008] Extracting tumor imaging features from the original dataset to obtain initial features, and sequentially performing zero importance analysis and correlation screening to obtain final screening features with high importance and low correlation;

[0009] Based on the final screening features and the annotated lesion type labels, a thyroid lymphoma prediction model is obtained by machine learning training;

[0010] Feature extraction is performed on the thyroid ultrasound segmentation image to be tested, which segments the suspected lesion area, to obtain the corresponding final screening features;

[0011] Based on the corresponding final screening characteristics, the thyroid lymphoma prediction model is used to predict thyroid lymphoma.

[0012] Furthermore, the zero importance analysis is performed on the initial features, including:

[0013] A random forest model is trained based on the original data set to calculate the importance score of each initial feature under the true label condition; wherein the true label is a lesion type label corresponding to the lesion area in the thyroid ultrasound segmentation image;

[0014] Randomly scrambling the lesion type labels to obtain a pseudo-label dataset, and retraining the random forest model based on the pseudo-label dataset to calculate the importance score of each initial feature under the pseudo-label condition;

[0015] Compare the distribution of the importance scores of each initial feature under the real label and pseudo label conditions to obtain the zero importance score of the initial feature;

[0016] Among all the initial features, features with zero importance scores greater than the preset zero threshold are screened out to obtain high-importance features.

[0017] Furthermore, the high-importance features are screened for relevance to obtain final screening features, including:

[0018] Calculate the correlation between each high-importance feature to obtain the correlation matrix;

[0019] sorting the high-importance features based on the zero-importance scores;

[0020] Traverse each high-importance feature in the sorting order and combine it with the correlation matrix, perform pairwise comparison, and screen out features whose correlation is lower than the preset correlation threshold to obtain the final screening features.

[0021] Furthermore, the Gini index is used to calculate the importance score of each initial feature, and the zero importance score of the initial feature is obtained based on the following formula:

[0022] ,

[0023] in, is the zero importance score of the initial feature i; is the importance score of the initial feature under the true label condition; is the importance score of the initial feature under the pseudo label condition; P() is the probability operation.

[0024] Furthermore, the main categories of the initial features include first-order features, shape features and high-level texture features. Based on the final screening features and the annotated lesion type labels, a thyroid lymphoma prediction model is obtained using random forest training.

[0025] Furthermore, automatic sampling technology was used to train a random forest model to obtain a thyroid lymphoma prediction model, including:

[0026] Based on the final screening features and the annotated lesion type labels, a training data set is obtained, and a plurality of sub-data sets are generated by sampling the training data set with replacement;

[0027] Based on each sub-dataset, the corresponding decision tree in the random forest-based thyroid lymphoma prediction model was trained independently;

[0028] The prediction results of each decision tree are integrated based on the following formula to obtain the final prediction result of thyroid lymphoma:

[0029] ,

[0030] in, To predict the final outcome of thyroid lymphoma; is the prediction result of the i-th decision tree; is the total number of decision trees.

[0031] Furthermore, when independently training the corresponding decision tree, the sample features and labels in the sub-dataset are sampled with replacement, and the corresponding training subset is generated based on the following formula:

[0032] ,

[0033] in, is the training subset corresponding to the i-th decision tree; Respectively represent the distribution of sample characteristics and label distribution The training samples are independently sampled from .

[0034] Furthermore, the tumor imaging features of the original dataset are extracted, including:

[0035] The grayscale histogram is used to extract the global intensity features of the thyroid ultrasound segmentation image in the original dataset and obtain the first-order features.

[0036] The geometric shape features are extracted using the area and perimeter of the segmented region to obtain the shape features;

[0037] Gray-level co-occurrence matrix, gray-level dependency matrix and gray-level run-length matrix are used to extract local structural information and obtain high-level texture features.

[0038] On the other hand, an embodiment of the present invention provides a thyroid lymphoma prediction device based on machine learning, comprising:

[0039] An image acquisition module, used for acquiring an ultrasound image containing a thyroid region;

[0040] a segmentation and annotation module, configured to segment the lesion area on the thyroid ultrasound image to generate a thyroid ultrasound segmented image; and further configured to annotate the lesion type label including lymphoma on the thyroid ultrasound segmented image;

[0041] a feature extraction module for extracting tumor imaging features from the thyroid ultrasound segmentation image to obtain initial features, and performing zero-importance analysis and correlation screening on the initial features in sequence to obtain final screening features with high importance and low correlation; and further for performing feature extraction on the thyroid ultrasound segmentation image to be tested that has segmented the suspected lesion area to obtain corresponding final screening features;

[0042] A prediction module, configured to obtain a thyroid lymphoma prediction model using machine learning training based on the final screening features and the annotated lesion type labels; and further configured to predict thyroid lymphoma using the thyroid lymphoma prediction model based on the corresponding final screening features;

[0043] The human-computer interaction module is used to output prediction results for the loaded images and feature data using visualization software and interface.

[0044] Furthermore, the feature extraction module is used to perform zero importance analysis on the initial features, including:

[0045] An original dataset is obtained based on the thyroid ultrasound segmentation image labeled with lesion type labels, a random forest model is trained based on the original dataset, and the importance score of each initial feature under the true label condition is calculated; wherein the true label is the lesion type label corresponding to the lesion area in the thyroid ultrasound segmentation image;

[0046] Randomly scrambling the lesion type labels to obtain a pseudo-label dataset, and retraining the random forest model based on the pseudo-label dataset to calculate the importance score of each initial feature under the pseudo-label condition;

[0047] Compare the distribution of the importance scores of each initial feature under the real label and pseudo label conditions to obtain the zero importance score of the initial feature;

[0048] Among all the initial features, features with zero importance scores greater than the preset zero threshold are screened out to obtain high-importance features.

[0049] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0050] 1. The present invention proposes a method for segmenting and labeling lesion areas based on thyroid ultrasound images. Feature extraction is performed by extracting tumor imaging category features, combining zero importance analysis and correlation analysis. The extracted highly representative and generalized tumor imaging features are used as input, and a thyroid lymphoma prediction model is obtained in combination with machine learning training. The corresponding final screening features are extracted from the thyroid ultrasound segmentation image to be tested, and the thyroid lymphoma prediction model is used to achieve accurate prediction of thyroid lymphoma.

[0051] 2. To compensate for the limitations of a single feature, first-order features, shape features, and high-level texture features are extracted from thyroid ultrasound segmentation images to improve the generalization ability of the model. The zero-importance analysis and correlation judgment of the random forest model are used to further reduce the dimensionality and remove redundancy of the extracted multiple features, and features with high label correlation are screened out, thereby improving the accuracy and efficiency of model prediction.

[0052] 3. The performance of the prediction models constructed by the multi-layer perceptron algorithm and the random forest algorithm was compared. The prediction model based on the random forest algorithm was selected to predict thyroid lymphoma. This improved the interpretability of the model. Combined with the redundant processing of input features, it reduced the computational cost brought by the random forest algorithm and increased the speed of model prediction. This provides clinicians with a powerful decision support tool to achieve early warning and precise prevention and control of thyroid lymphoma.

[0053] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0055] Figure 1 This is a flowchart of a method for predicting thyroid lymphoma based on machine learning according to an embodiment of the present invention;

[0056] Figure 2 This is a heat map of high-importance and low-correlation features obtained during the feature extraction phase of an embodiment of the present invention;

[0057] Figure 3 Schematic diagram of the principle of the random forest algorithm according to an embodiment of the present invention;

[0058] Figure 4 This is a schematic diagram of the multi-layer perceptron algorithm principle according to an embodiment of the present invention;

[0059] Figure 5 This is a confusion matrix result diagram obtained when processing the random forest prediction model according to an embodiment of the present invention;

[0060] Figure 6 This is a confusion matrix result diagram obtained when processing the multi-layer perceptron prediction model according to an embodiment of the present invention;

[0061] Figure 7 This is a ROC curve diagram obtained when the random forest prediction model according to the embodiment of the present invention is processed;

[0062] Figure 8 This is a ROC curve diagram obtained when processing the multi-layer perceptron prediction model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0064] Example 1

[0065] A specific embodiment of the present invention discloses a method for predicting thyroid lymphoma based on machine learning, such as Figure 1 As shown, the following steps are included:

[0066] Step S1: annotating the thyroid ultrasound image with the lesion area segmented with a lesion type label including lymphoma to obtain an original data set;

[0067] Step S2: extracting tumor imaging features from the original dataset to obtain initial features, and sequentially performing zero importance analysis and correlation screening to obtain final screening features with high importance and low correlation;

[0068] Step S3: Based on the final screening features and the annotated lesion type labels, a thyroid lymphoma prediction model is obtained by machine learning training;

[0069] Step S4, performing feature extraction on the thyroid ultrasound segmentation image to be tested, in which the suspected lesion area is segmented, to obtain corresponding final screening features;

[0070] Step S5: Based on the corresponding final screening features, the thyroid lymphoma prediction model is used to predict thyroid lymphoma.

[0071] Through the above method, tumor imaging features are extracted based on thyroid ultrasound segmentation images with annotated lesion type labels, zero importance analysis and correlation analysis are performed on the extracted multiple features, and a thyroid lymphoma prediction model is obtained based on feature training after redundancy removal. The model is used to predict the thyroid ultrasound segmentation images to be tested, achieving high precision and high robustness in thyroid lymphoma prediction.

[0072] Specifically, in step S1, the lesion regions are divided and lesion type labels are annotated on the acquired ultrasound images containing thyroid lymphoma regions. The specific steps include:

[0073] S11. Screening of qualified thyroid ultrasound images: Selecting thyroid ultrasound images with high clarity, resolution that meets the requirements, and that meet a certain number of conditions;

[0074] S12. Determine the preliminary extent of the lymphoma lesion area on the thyroid ultrasound image: Combine the thyroid ultrasound image with the diagnosis of a professional physician to determine the approximate extent of the lymphoma area, and verify its accuracy with a professional physician;

[0075] S13, segmenting the region of the lesion type such as lymphoma in the thyroid ultrasound image to generate a corresponding thyroid ultrasound segmentation image;

[0076] For example, based on real-world clinical data, we obtained basic case information and corresponding thyroid ultrasound images, including 32 cases of primary thyroid malignant lymphoma (PTML). A control group included four cases of papillary thyroid carcinoma, eight cases of follicular thyroid tumors, and eight cases of other thyroid diseases, totaling 52 cases and 145 sets of thyroid ultrasound images. Using the medical image processing software 3D-Slicer, we mapped the lesion area on the thyroid ultrasound image to obtain a segmented thyroid ultrasound image.

[0077] S14. Manually label the lesion type labels on the thyroid ultrasound segmentation images with segmented lesion areas, save them, and review and adjust them by professional physicians to finally obtain labeled segmentation images. The labeled thyroid ultrasound segmentation images constitute the original dataset.

[0078] Specifically, in step S2, before feature extraction, the thyroid ultrasound segmented images in the original dataset are preprocessed by scaling, normalization, and denoising. In order to more comprehensively capture the complex information related to thyroid lymphoma, the specific steps for extracting tumor imaging features based on the preprocessed images and obtaining the final screening features are as follows:

[0079] S21. Extract multiple types of tumor imaging features based on the original data set to obtain initial features, including main category features of first-order features, shape features, and high-level texture features; wherein each main category contains multiple subcategory features, that is, the initial features contain tumor imaging features of multiple subcategories.

[0080] For example, the feature extraction process of the third-party Python pyradiomics package library is performed, and multiple image types and optional features are enabled to obtain tumor imaging features of thyroid ultrasound segmentation images as follows:

[0081] S211. According to the requirements of thyroid lymphoma segmentation and the characteristics of thyroid ultrasound images, characteristic parameters for extracting tumor imaging features are set, including the width of the histogram bin, the Gaussian filter standard deviation sigma, the interpolator type, the pixel spacing, and the voxel array offset;

[0082] The interpolator type is adjusted to scale or rotate the segmented image to ensure smoothness and detail preservation. The Gaussian filter standard deviation (sigma) is adjusted to smooth the image and remove noise while preserving the image's primary structural features, achieving image preprocessing. During feature extraction, the histogram width parameter is adjusted to set the image's grayscale accuracy for texture analysis. The pixel spacing is adjusted to control the image's spatial resolution for shape and texture analysis.

[0083] S212. Using grayscale histogram features, mean, standard deviation, etc., extract the global intensity features of the thyroid ultrasound segmentation image in the original data set to obtain first-order features; using the area, perimeter, shape factor, etc. of the segmented region, extract the geometric shape of the lesion tumor region to obtain shape features; using the gray-level co-occurrence matrix (GLCM), gray-level dependency matrix (GLDM), and gray-level run-length matrix (GLRLM) to extract local structural information of the image to obtain high-level texture features;

[0084] Through the above approach, all subcategories corresponding to first-order features, shape features, and high-level texture features in tumor imaging were activated, resulting in the extraction of 1,042 tumor imaging features. By comprehensively considering multiple features and leveraging their combined effect, diagnostic accuracy was significantly improved, addressing the limitation that a single feature cannot provide sufficient information to ensure accurate disease diagnosis.

[0085] S22. Use principal component analysis (PCA) technology to reduce the dimension of the feature space and select the feature subset that has the greatest contribution to the classification task, thereby reducing computational complexity and preventing overfitting.

[0086] Specifically, based on the extracted initial features of the thyroid ultrasound segmentation image, zero importance analysis is performed to obtain high importance features. The specific steps are as follows:

[0087] S221. Train a random forest model based on the original dataset, calculate the feature importance score of each initial feature under the true label condition, and record the feature importance distribution;

[0088] The true label is the lesion type label corresponding to the lesion area in the thyroid ultrasound segmentation image; after all initial features are arranged according to the importance scores, the feature importance distribution is obtained.

[0089] For example, the Gini index is used to calculate the importance score of each initial feature. The purity of the initial features is measured by the Gini index, which can best segment all the initial features. The calculation formula is:

[0090] ,

[0091] in, is the importance score of the initial feature i; is the sample probability corresponding to the initial feature i; is the total number of initial feature categories.

[0092] The lower the Gini index, the lower the feature importance score under true label conditions, indicating a purer feature. The higher the Gini index, the more mixed the feature category distribution. By comparing the impact of different feature Gini indices, the decision tree model can automatically select the most useful features for data partitioning, thereby optimizing classification results.

[0093] S222, performing a random scrambling operation on the lesion type labels to destroy the true mapping relationship between features and labels, constructing a pseudo-label data set, and ensuring that the pseudo-label data does not contain valid feature information;

[0094] S223. Retrain the random forest model 100 times based on the pseudo-label dataset, calculate the importance score of each initial feature under the pseudo-label condition, and record the feature importance distribution under the pseudo-label condition;

[0095] S224. Compare the distribution of the importance score of each feature under the real label condition and the pseudo label condition, and calculate the zero importance score of the initial feature;

[0096] For example, for each feature , the calculation formula of zero importance score is:

[0097] ,

[0098] in, is the zero importance score of the initial feature i, that is, the cumulative probability of the actual importance value in the random distribution; is the importance score of the initial feature under the true label condition; is the importance score of the initial feature under pseudo-label conditions; P() is the probability operation. A higher score indicates a greater actual contribution of the feature to the classification task.

[0099] It should be noted that numerical simulation methods are typically used to calculate probabilities using statistics to obtain the zero importance score. A large number of samples (e.g., N) are drawn from the pseudo-labeled results for a feature. The number of samples whose importance scores are less than the corresponding "actual importance" (n) is then calculated as score = n / N. For example, if 10,000 samples are drawn from the null hypothesis distribution, and 8,000 of them are less than the actual importance value, then the score = 8,000 / 10,000 = 0.8.

[0100] S225. Screen out features whose zero importance scores are greater than a preset zero threshold from all initial features to obtain high-importance features.

[0101] For example, the zeroing threshold is set to 0.6, and based on the zero importance analysis results, features with significant importance are screened out to optimize the feature set.

[0102] S23. Use correlation screening to select features that are highly correlated with the lesion type label and remove redundant features from high-importance features, including:

[0103] S231. Calculate the correlation between each high-importance feature using the Pearson correlation coefficient to obtain a correlation matrix;

[0104] S232. Sort the high-importance features based on the zero-importance score;

[0105] S233. Traverse each high-importance feature in the sorting order, combine with the correlation matrix, perform pairwise comparison, screen out features with correlation coefficients lower than the preset correlation threshold, and obtain final screening features with low correlation and high zero importance score.

[0106] Exemplarily, high-importance features are arranged from high to low according to the corresponding zero-importance scores to form a zero-importance list, and each feature in the zero-importance list is selected from high to low for pairwise comparison with other high-importance features in the list. First, the first feature in the zero-importance list is taken as the selected feature, and the correlation coefficient between the second feature in the zero-importance list and the selected feature is determined. If the correlation is higher than the threshold value of 0.6, it means that the second feature is highly correlated with the selected feature, and the second feature is deleted from the zero-importance list; otherwise, the second feature is added to the retained feature list; then, the correlation between the first feature and the third feature in the zero-importance list is determined; similarly, after successively determining the correlation between the first feature and all other features in the zero-importance list, the second feature retained in the zero-importance list is taken as the selected feature, and the correlation between the currently selected feature and the remaining features is determined in turn. After zero-importance analysis and correlation analysis, 17 features with low correlation and high importance are obtained in the retained feature list, and their heat map is shown as follows. Figure 2 As shown in the figure, a more effective feature set is finally obtained, which is the final screening feature containing three categories: first-order features, shape features and high-level texture features.

[0107] Through the above method, the preliminary features are selected and optimized, the most informative tumor imaging features are selected, and the features most relevant to diagnosis are retained to optimize the performance of the model and reduce computational complexity.

[0108] Specifically, in step S3, based on the final screening features and the annotated lesion type labels, a thyroid lymphoma prediction model is obtained using random forest training.

[0109] For example, the network structure of the thyroid lymphoma prediction model is obtained based on random forest algorithm training, such as Figure 3As shown in the figure, an efficient thyroid lymphoma prediction model was constructed by introducing the random forest algorithm. This model integrates multiple decision trees and multidimensional features in the data, focusing on the core factors of lymphoma, including:

[0110] S31. Build the basic structure of the random forest model and set the decision tree parameters;

[0111] S32. Use the Bootstrap Sampling (Bagging) module to reduce the variance of the random forest model and avoid the overfitting problem of a single decision tree. The specific steps include the following:

[0112] S321. Based on the final screening features and the annotated lesion type labels, a training data set is obtained. The training data set is sampled with replacement to generate multiple sub-data sets. Each sub-data set is obtained by:

[0113] ,

[0114] in, is the i-th sub-dataset; Indicates the number of samples in the i-th sub-dataset; Represent the features and labels of the samples respectively.

[0115] It should be noted that the thyroid lymphoma prediction model is a binary classification model. The labels of the input final screening features are adjusted, and the final screening features of patients with thyroid lymphoma are marked as 1, and the rest are marked as 0.

[0116] For example, the label of the ultrasound image of patients with thyroid lymphoma is adjusted to 1, that is, the label corresponding to the final screening feature obtained thereby is adjusted to 1; the label of the ultrasound image of other thyroid diseases in the control group is adjusted to 0, that is, the label corresponding to the final screening feature obtained thereby is adjusted to 0; finally, 76 labels of 1 and 69 labels of 0 are obtained, and these 145 groups of thyroid ultrasound segmentation images are divided into 80% training data sets and 20% validation sets.

[0117] S322. Based on each sub-dataset, independently training the corresponding decision tree in the random forest-based thyroid lymphoma prediction model;

[0118] For example, the training process of each decision tree is set as follows:

[0119] ,

[0120] in, Represents the i-th decision tree, through the sub-dataset Conduct training.

[0121] Furthermore, during the training process of each tree, the sample features and labels in the subset data set are sampled with replacement. The generation process of the training subset is expressed as:

[0122]

[0123] in, is the training subset corresponding to the i-th decision tree; Respectively represent the distribution of sample characteristics and label distribution The training samples are independently sampled from .

[0124] S323. Integrate the prediction results of each decision tree based on the following formula to obtain the final prediction result of thyroid lymphoma:

[0125] ,

[0126] in, To predict the final outcome of thyroid lymphoma; is the prediction result of the i-th decision tree; is the total number of decision trees.

[0127] In order to compare the effects of different machine learning models on thyroid lymphoma prediction, as another optional solution, a thyroid lymphoma prediction model was obtained by training a multi-layer perceptron based on the final screening features and the annotated lesion type labels.

[0128] For example, 80% of the 145 relabeled data sets are used as training sets and 20% as test sets. The network structure of the thyroid lymphoma prediction model based on the multi-layer perceptron (MLP) algorithm is as follows: Figure 4 As shown in Figure 2, the model performs deep learning on features through a multi-layer fully connected network, integrating nonlinear relationships at different levels.

[0129] When training the above two prediction models, confusion matrices are introduced to evaluate the model performance, and the training parameters are continuously adjusted to make the training results stable. The confusion matrix results are as follows: Figure 5 and Figure 6 shown.

[0130] The receiver operating characteristic (ROC) curve was used to evaluate the two thyroid lymphoma prediction models obtained by machine learning. Figure 7 、 Figure 8As shown in the figure, the random forest model achieved an accuracy of 0.90, with an average accuracy of 0.84 for five-fold cross-validation and 0.81 for ten-fold cross-validation. The area under the receiver operating characteristic (ROC) curve (AUC) reached 0.90. The multilayer perceptron model achieved an accuracy of 0.86, with an AUC of 0.90.

[0131] The prediction results are clearly shown in Table 1, demonstrating that both models demonstrate excellent performance. However, the random forest model effectively handles high-dimensional data, avoids overfitting, and is more interpretable than the multilayer perceptron, making it the preferred approach. The random feature selection of the bagging module reduces correlations between trees, ensuring high accuracy and robustness in predicting thyroid lymphoma prevalence. Furthermore, pre-processing of the multidimensional input features for redundancy further enhances the model's efficiency, providing strong support for developing personalized treatment strategies.

[0132] Table 1

[0133]

[0134] Specifically, in step S4, the suspected lesion area is segmented in the thyroid ultrasound image to be tested to obtain a corresponding thyroid ultrasound segmentation image, and then based on the category of the final screening feature, the corresponding features are specifically extracted to obtain the corresponding final screening feature.

[0135] Specifically, in step S5, the corresponding final screening features obtained are input into the thyroid lymphoma prediction model to obtain a prediction result of the thyroid lymphoma prevalence; the prediction result is a binary classification output, if the patient has thyroid lymphoma, the prediction result is 1, otherwise it is 0.

[0136] Compared with the existing technology, the present embodiment provides a thyroid lymphoma prediction method based on machine learning. By extracting multiple tumor imaging features from labeled thyroid ultrasound segmentation images, using zero importance analysis and correlation screening to remove redundancy to obtain the final screening features, a thyroid lymphoma prediction model is obtained based on machine learning training. The prediction model is used on the thyroid ultrasound segmentation image to be tested, and ultimately an accurate prediction of thyroid lymphoma is achieved. On the one hand, the extraction of multiple types of features including first-order, shape, and high-level texture features can comprehensively reflect the various levels of features of thyroid lymphoma, provide rich input information for the prediction model, and then improve the accuracy of the model through de-redundancy processing; on the other hand, the prediction is made using a random model based on dimensionality reduction and de-redundant features, which provides model interpretability while reducing computational costs and improving model efficiency.

[0137] Example 2

[0138] Another specific embodiment of the present invention discloses a thyroid lymphoma prediction device based on machine learning, comprising:

[0139] An image acquisition module, used for acquiring a plurality of ultrasound images including the thyroid region;

[0140] a segmentation and annotation module, configured to segment the lesion area on the thyroid ultrasound image to generate a thyroid ultrasound segmented image; and further configured to annotate the lesion type label including lymphoma on the thyroid ultrasound segmented image;

[0141] a feature extraction module for extracting tumor imaging features from the thyroid ultrasound segmentation image to obtain initial features, and performing zero-importance analysis and correlation screening on the initial features in sequence to obtain final screening features with high importance and low correlation; and further for performing feature extraction on the thyroid ultrasound segmentation image to be tested that has segmented the suspected lesion area to obtain corresponding final screening features;

[0142] A prediction module, configured to obtain a thyroid lymphoma prediction model using machine learning training based on the final screening features and the annotated lesion type labels; and further configured to predict thyroid lymphoma using the thyroid lymphoma prediction model based on the corresponding final screening features;

[0143] The human-computer interaction module is used to output prediction results for the loaded images and feature data using visualization software and interface.

[0144] Furthermore, the feature extraction module is used to perform zero importance analysis on the initial features, including:

[0145] An original dataset is obtained based on the thyroid ultrasound segmentation image labeled with lesion type labels, a random forest model is trained based on the original dataset, and the importance score of each initial feature under the true label condition is calculated; wherein the true label is the lesion type label corresponding to the lesion area in the thyroid ultrasound segmentation image;

[0146] Randomly scrambling the lesion type labels to obtain a pseudo-label dataset, and retraining the random forest model based on the pseudo-label dataset to calculate the importance score of each initial feature under the pseudo-label condition;

[0147] Compare the distribution of the importance scores of each initial feature under the real label and pseudo label conditions to obtain the zero importance score of the initial feature;

[0148] Among all the initial features, features with zero importance scores greater than the preset zero threshold are screened out to obtain high-importance features.

[0149] The system can predict thyroid lymphoma according to the thyroid lymphoma prediction method based on machine learning described in any one of the schemes in Example 1. The relevant parts are referenced from each other and are not repeated in this embodiment.

[0150] Compared with the existing technology, this embodiment provides a thyroid lymphoma prediction device based on machine learning. Based on the image acquisition module, segmentation and annotation module, feature extraction module, prediction module and human-computer interaction module, a thyroid lymphoma prediction device based on machine learning is proposed to realize the prediction of thyroid lymphoma, provide clinicians with a powerful decision support tool, and achieve early warning and precise prevention and control of thyroid lymphoma.

[0151] Example 3

[0152] Another specific embodiment of the present invention discloses a computing device for predicting thyroid lymphoma based on machine learning, comprising:

[0153] a memory configured to store computer-executable instructions for machine learning-based prediction of thyroid lymphoma;

[0154] A processor is configured to execute the computer-executable instructions and process the thyroid lymphoma prediction method described in any one of the schemes in Example 1.

[0155] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0156] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A thyroid lymphoma prediction method based on machine learning, characterized in that: The steps include: The thyroid ultrasound images with segmented lesion areas are labeled with lesion type labels including lymphoma to obtain the original dataset; Extracting tumor imaging features from the original dataset to obtain initial features, and sequentially performing zero importance analysis and correlation screening to obtain final screening features with high importance and low correlation; Based on the final screening features and the annotated lesion type labels, a thyroid lymphoma prediction model is obtained by machine learning training; Feature extraction is performed on the thyroid ultrasound segmentation image to be tested, which segments the suspected lesion area, to obtain the corresponding final screening features; Based on the corresponding final screening characteristics, the thyroid lymphoma prediction model is used to predict thyroid lymphoma; Perform zero importance analysis on the initial features, including: A random forest model is trained based on the original data set to calculate the importance score of each initial feature under the true label condition; wherein the true label is a lesion type label corresponding to the lesion area in the thyroid ultrasound segmentation image; Randomly scrambling the lesion type labels to obtain a pseudo-label dataset, and retraining the random forest model based on the pseudo-label dataset to calculate the importance score of each initial feature under the pseudo-label condition; Compare the distribution of the importance scores of each initial feature under the real label and pseudo label conditions to obtain the zero importance score of the initial feature; Among all the initial features, features with zero importance scores greater than the preset zero threshold are screened out to obtain high-importance features.

2. A thyroid lymphoma prediction method based on machine learning according to claim 1, characterized in that: The high-importance features are screened for relevance to obtain final screening features, including: Calculate the correlation between each high-importance feature to obtain the correlation matrix; sorting the high-importance features based on the zero-importance scores; Traverse each high-importance feature in the sorting order and combine it with the correlation matrix, perform pairwise comparison, and screen out features whose correlation is lower than the preset correlation threshold to obtain the final screening features.

3. A thyroid lymphoma prediction method based on machine learning according to claim 1, characterized in that: The Gini index is used to calculate the importance score of each initial feature, and the zero importance score of the initial feature is obtained based on the following formula: ， in, is the zero importance score of the initial feature i; is the importance score of the initial feature under the true label condition; is the importance score of the initial feature under the pseudo label condition; P() is the probability operation.

4. The method for predicting thyroid lymphoma based on machine learning according to claim 1, wherein: The main categories of the initial features include first-order features, shape features and high-level texture features. Based on the final screening features and the annotated lesion type labels, a thyroid lymphoma prediction model is obtained using random forest training.

5. The method for predicting thyroid lymphoma based on machine learning according to claim 4, wherein: The thyroid lymphoma prediction model was obtained by using automatic sampling technology and random forest training, including: Based on the final screening features and the annotated lesion type labels, a training data set is obtained, and a plurality of sub-data sets are generated by sampling the training data set with replacement; Based on each sub-dataset, the corresponding decision tree in the random forest-based thyroid lymphoma prediction model was trained independently; The prediction results of each decision tree are integrated based on the following formula to obtain the final prediction result of thyroid lymphoma: ， in, To predict the final outcome of thyroid lymphoma; is the prediction result of the i-th decision tree; is the total number of decision trees.

6. The method for predicting thyroid lymphoma based on machine learning according to claim 5, wherein: When independently training the corresponding decision tree, the sample features and labels in the sub-dataset are sampled with replacement, and the corresponding training subset is generated based on the following formula: ， in, is the training subset corresponding to the i-th decision tree; Respectively represent the distribution of sample characteristics and label distribution The training samples are independently sampled from .

7. A method for predicting thyroid lymphoma based on machine learning according to any one of claims 1 to 6, characterized in that: Extracting tumor imaging features from the original dataset includes: The grayscale histogram is used to extract the global intensity features of the thyroid ultrasound segmentation image in the original dataset and obtain the first-order features. The geometric shape features are extracted using the area and perimeter of the segmented region to obtain the shape features; Gray-level co-occurrence matrix, gray-level dependency matrix and gray-level run-length matrix are used to extract local structural information and obtain high-level texture features.

8. A thyroid lymphoma prediction device based on machine learning, characterized in that: include: An image acquisition module, used for acquiring an ultrasound image containing a thyroid region; a segmentation and annotation module, configured to segment the lesion area on the thyroid ultrasound image to generate a thyroid ultrasound segmented image; and further configured to annotate the lesion type label including lymphoma on the thyroid ultrasound segmented image; a feature extraction module, configured to extract tumor imaging features from the thyroid ultrasound segmentation image to obtain initial features, and sequentially perform zero importance analysis and correlation screening on the initial features to obtain final screening features with high importance and low correlation; It is also used to extract features from the thyroid ultrasound segmentation image to be tested that segments the suspected lesion area to obtain the corresponding final screening features; A prediction module, configured to obtain a thyroid lymphoma prediction model using machine learning training based on the final screening features and the annotated lesion type labels; and further configured to predict thyroid lymphoma using the thyroid lymphoma prediction model based on the corresponding final screening features; The human-computer interaction module is used to output prediction results based on the loaded images and feature data using visualization software and interfaces; Performing zero importance analysis on the initial features using the feature extraction module includes: An original dataset is obtained based on the thyroid ultrasound segmentation image labeled with lesion type labels, a random forest model is trained based on the original dataset, and the importance score of each initial feature under the true label condition is calculated; wherein the true label is the lesion type label corresponding to the lesion area in the thyroid ultrasound segmentation image; Randomly scrambling the lesion type labels to obtain a pseudo-label dataset, and retraining the random forest model based on the pseudo-label dataset to calculate the importance score of each initial feature under the pseudo-label condition; Compare the distribution of the importance scores of each initial feature under the real label and pseudo label conditions to obtain the zero importance score of the initial feature; Among all the initial features, features with zero importance scores greater than the preset zero threshold are screened out to obtain high-importance features.

Images