An imaging-based method, system and device for predicting EPH after parathyroidectomy

Abstract

This application relates to the field of intelligent medicine, and particularly relates to a method, system and device for predicting EPH after parathyroidectomy based on images. It includes obtaining parathyroid images of a patient; extracting EPH highly correlated features based on the parathyroid images to predict EPH after surgery to obtain a prediction result; wherein, the construction process of the EPH highly correlated features is as follows: S1: obtaining a parathyroid image set, EPH grouping labels, and blood calcium concentration labels of the patient; S2: extracting primary features from the parathyroid image set based on the EPH grouping labels and blood calcium concentration labels; S3: performing feature screening on the primary features to obtain the EPH highly correlated features. This method predicts whether EPH occurs after parathyroidectomy for patients with secondary hyperparathyroidism through radiomics features, and has good clinical value.

Description

Technical Field

[0001] The present application relates to the field of intelligent medicine, and particularly relates to a method, a system, a device and a computer-readable storage medium for predicting EPH after parathyroidectomy based on images. Background Art

[0002] Chronic kidney disease (CKD) is one of the chronic diseases that seriously threaten the health of the people in China and has also become a serious medical burden in China. Secondary hyperparathyroidism (SHPT) is one of the most common complications in CKD hemodialysis patients, mainly manifested as elevated parathyroid hormone, calcium and phosphorus metabolism disorders, ectopic calcification, bone pain, osteoporosis, etc. At present, the treatment of SHPT mainly includes medical conservative treatment and surgical resection. Medical conservative treatment (calcium mimetics) has obvious effects in the initial stage of the disease, but with the progression of the disease, a series of symptoms such as uncontrollable hypercalcemia and hyperphosphatemia will appear, seriously affecting the prognosis of patients. 10% - 30% of end-stage renal disease patients ultimately require parathyroidectomy (PTX). In addition, for CKD patients who are difficult to treat with drugs or cannot afford expensive drugs, surgical treatment is still the best choice to control the level of parathyroid hormone. However, early postoperative hypocalcemia (EPH), which is extremely likely to occur due to the sudden decrease of parathyroid hormone after surgery, is also the most core issue surrounding the perioperative management. Previous literature reports that the incidence of EPH after PTX is very high, ranging from about 30% to 80%, and mild clinical symptoms may be accompanied by weakness, headache, paresthesia, intestinal obstruction, malabsorption and muscle spasm, while severe hypocalcemia can lead to sudden death. However, the current guidelines have not yet formed a definite conclusion, and there is a lack of a reliable preoperative risk scoring method. In addition, the risk factors reported in previous literature mainly focus on preoperative serum laboratory indexes, and the conclusions are inconsistent, and the dynamic fluctuations of biochemical indexes brought about by dialysis in CKD patients are ignored, which restricts the stability of its prediction. Summary of the Invention

[0003] In view of the above problems, the present invention discovers through experiments that the self-calcification phenomenon of preoperative parathyroid lesions in SHPT patients can independently predict the occurrence of EPH after PTX. However, in the previous research, it was also found that the manifestation forms of parathyroid lesion calcification are diverse, the number of involved lesions and the distribution location of calcification are highly heterogeneous, and there are also microcalcification foci that cannot be judged by the naked eye in CT, etc. This also indicates that relevant imaging information should be further deeply explored. Therefore, the present invention proposes a method for predicting EPH after parathyroidectomy based on images, including:

[0004] Obtain the parathyroid image of the patient;

[0005] Extract EPH highly correlated features based on the parathyroid image to predict EPH after surgery to obtain a prediction result;

[0006] Among them, the construction process of the EPH highly correlated features is as follows:

[0007] S1: Obtain the parathyroid image set, EPH grouping label, and blood calcium concentration label of the patient;

[0008] S2: Extract features from the parathyroid image set based on the EPH grouping label and blood calcium concentration label to obtain primary features;

[0009] S3: Screen the primary features to obtain the EPH highly correlated features.

[0010] Furthermore, the specific content of S2 is as follows:

[0011] S21: Extract the first radiomics features from the parathyroid image set based on the EPH grouping label; extract primary features from the first radiomics features based on the blood calcium concentration label; the blood calcium concentration label is a binary classification label, and is divided into the first blood calcium concentration and the second blood calcium concentration through a preset blood calcium concentration threshold;

[0012] Optionally, S21 is replaced by S22: Extract the first radiomics features from the parathyroid image set based on the EPH grouping label; at the same time, extract the second radiomics features from the parathyroid image set based on the blood calcium concentration label; perform an intersection operation on the first radiomics features and the second radiomics features to obtain primary features;

[0013] Optionally, S21 is replaced by S23: Extract the second radiomics features from the parathyroid image set based on the blood calcium concentration label; extract primary features from the second radiomics features based on the EPH grouping label.

[0014] Or the specific process of S2 is as follows:

[0015] S24: Extract the first radiomics features from the parathyroid image set based on the EPH grouping label; calculate the correlation between the first radiomics features and the blood calcium concentration label to obtain primary features; the blood calcium concentration label is a continuous variable label;

[0016] Optionally, S24 is replaced by S25: The parathyroid image set is subjected to feature extraction based on the EPH grouping label to obtain the first radiomics feature; meanwhile, the parathyroid image set is subjected to feature extraction based on the blood calcium concentration label to obtain the third radiomics feature; the correlation between the first radiomics feature and the third radiomics feature is calculated to obtain the primary feature;

[0017] Optionally, S24 is replaced by S26: The parathyroid image set is subjected to feature extraction based on the blood calcium concentration label to obtain the third radiomics feature, and the correlation between the third radiomics feature and the EPH grouping label is calculated to obtain the primary feature.

[0018] Further, the steps of S3 are as follows:

[0019] The first step: Input the primary feature into a machine learning model to preliminarily sort and screen the features to obtain candidate features;

[0020] The second step: Analyze the contribution degree of the candidate features through SHAP values to obtain contribution degree features;

[0021] The third step: Screen out and sort the irrelevant features of the contribution degree features to obtain EPH highly correlated features.

[0022] The highly correlated features include one or more of the following: 3D m2 local binary pattern gray level size matrix gray non-uniformity normalization, 3D k local binary pattern gray level size matrix region entropy, wavelet LLL first-order skewness, 3D k local binary pattern gray level run matrix run length non-uniformity normalization, wavelet HLL first-order skewness.

[0023] The method further includes obtaining clinical features. First, it is divided into EPH group and non-EPH group data based on the EPH grouping label and univariate analysis is performed to obtain statistically significant features, and then multivariate analysis is performed on the statistically significant features of the two groups to obtain clinical features. The clinical features are fused with the EPH highly correlated features and then postoperative EPH prediction is performed;

[0024] Optionally, the clinical features include one or more of the following: parathyroid calcification, hemoglobin, alkaline phosphatase, preoperative parathyroid hormone.

[0025] The method further includes multi-modal data feature prediction. First, patient clinical information and biochemical indexes are obtained and feature extraction is performed to obtain clinical information features and biochemical features. Postoperative EPH prediction is performed based on the clinical information features, biochemical features and the EPH highly correlated features to obtain a prediction result.

[0026] The object of the present invention is to provide an image-based EPH prediction system for parathyroidectomy, including:

[0027] When the system executes, it implements the above-mentioned image-based EPH prediction method after parathyroidectomy.

[0028] The purpose of the present invention is to provide an image-based EPH prediction system after parathyroidectomy, including:

[0029] An acquisition unit: to acquire parathyroid images of a patient;

[0030] A prediction unit: based on the parathyroid images, extract EPH highly correlated features to perform postoperative EPH prediction to obtain a prediction result;

[0031] Among them, the construction process of the EPH highly correlated features is as follows:

[0032] S1: Obtain parathyroid image sets and blood calcium concentration data sets of EPH patients and non-EPH patients;

[0033] S2: Based on the parathyroid image sets and blood calcium concentration data sets, perform feature extraction to obtain primary features of the EPH group and the non-EPH group;

[0034] S3: Perform feature screening on the primary features of the EPH group and the non-EPH group to obtain the EPH highly correlated features.

[0035] The purpose of the present invention is to provide an image-based EPH prediction device after parathyroidectomy, including:

[0036] A memory and a processor, the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, it implements the above-mentioned image-based EPH prediction method after parathyroidectomy.

[0037] The purpose of the present invention is to provide a computer-readable storage medium with a computer program thereon, including:

[0038] When the computer program is executed by a processor, it implements the above-mentioned image-based EPH prediction method after parathyroidectomy.

[0039] Advantages of the present invention:

[0040] 1. For the calcification phenomenon of parathyroid lesions, the image can provide information that doctors cannot subjectively judge and describe the heterogeneity of the lesions through quantification, providing a new method for predicting whether EPH occurs after parathyroidectomy, and performing EPH prediction through parathyroid radiomics features;

[0041] 2. By screening radiomics features highly correlated with EPH for postoperative EPH prediction, the accuracy and reliability of the prediction are improved;

[0042] 3. Feature screening is carried out using traditional radiomics feature screening and new radiomics screening methods; prediction model training is performed based on the data obtained from the screening methods, and a prediction model with excellent prediction effect is obtained;

[0043] 4. Radiomics features highly correlated with EPH, clinical information, and biochemical indicators are used to predict postoperative EPH, and the prediction model is trained for feature representation from multiple dimensions. Brief Description of the Drawings

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0045] Figure 1 It is a schematic flowchart of a method for predicting EPH after parathyroidectomy based on images provided by an embodiment of the present invention;

[0046] Figure 2 It is a schematic diagram of a system for predicting EPH after parathyroidectomy based on images provided by an embodiment of the present invention;

[0047] Figure 3 It is a schematic diagram of a device for predicting EPH after parathyroidectomy based on images provided by an embodiment of the present invention;

[0048] Figure 4 It is an image segmentation map of parathyroid lesions provided by an embodiment of the present invention;

[0049] Figure 5 It is a LASSO result diagram after feature screening by the traditional method provided by an embodiment of the present invention;

[0050] Figure 6 It is the shape values of 20 important features screened by the new screening method provided by an embodiment of the present invention;

[0051] Figure 7 It is a LASSO result diagram after feature screening by the new method provided by an embodiment of the present invention;

[0052] Figure 8 It is a ROC index result diagram of the model under each feature training set provided by an embodiment of the present invention;

[0053] Figure 9 It is a ROC index result diagram of the model under each feature test set provided by an embodiment of the present invention. Detailed Embodiments

[0054] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0055] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The sequence numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0056] Figure 1 A schematic diagram of an image-based method for predicting EPH after parathyroidectomy provided by an embodiment of the present invention specifically includes:

[0057] S101: Obtain the patient's parathyroid gland images;

[0058] In one embodiment, the parathyroid gland is one of the endocrine glands in the human body. The human body has two pairs of parathyroid glands, which are brown-yellow and shaped like soybeans, respectively located in the middle and lower parts of the back of the left and right lobes of the thyroid gland (or buried therein). The main function is to secrete parathyroid hormone (PTH for short) to regulate the metabolism of calcium and phosphorus in the body. If the parathyroid gland is hypofunctioning or completely removed (such as accidentally removed during thyroid surgery), PTH secretion is insufficient, causing blood calcium to gradually decrease, while blood phosphorus to gradually increase, leading to hypocalcemic convulsions and even death. Supplementing PTH and calcium salts can temporarily relieve symptoms. When the parathyroid gland is hyperfunctioning and PTH is secreted excessively, bone calcium enters the blood, and the kidney reabsorption of calcium is strengthened, while vitamin D3 is activated to become active D3, promoting the absorption of calcium by the small intestine, causing excessive blood calcium, and inhibiting the kidney reabsorption of phosphate, promoting the excretion of phosphorus in the urine, causing low blood phosphorus, thereby causing abnormal deposition of calcium salts in some tissues.

[0059] In one embodiment, the image includes one or more of the following: CT, ultrasound, MRI, and X-ray.

[0060] In a specific embodiment, uremic patients undergo PTX surgery under general anesthesia, and perioperative management is fraught with many challenges. For the special vulnerable group of uremic patients who "become poor due to illness" and "get sick due to illness", how to reduce postoperative complications and mortality is what surgeons should focus on, and it is also the key to improving the quality of life of this vulnerable group. However, due to the corresponding pathophysiological changes of their own diseases, postoperative hypocalcemia is the biggest adverse event after PTX. Preventing the occurrence of postoperative hypocalcemia can not only shorten the time of intravenous calcium supplementation, but also reduce the hospitalization cost and the risk of other postoperative complications.

[0061] After surgical treatment of patients with secondary hyperparathyroidism (SHPT), the incidence of severe hypocalcemia in the early postoperative period can be as high as 30 - 80%. If severe hypocalcemia is not corrected in a timely manner after surgery, it will not only prolong the patient's hospital stay, increase the frequency of blood sampling to monitor blood calcium levels, but also increase the hospitalization cost of uremic patients, and it is more likely to occur other serious complications or even death. At present, there is no exact preoperative prediction system internationally. The present invention focuses on solving the above clinical problems, which will help to strengthen the precise management after PTX in SHPT patients and reduce the risk of adverse events after SHPT.

[0062] In a specific embodiment, the cases included in the present invention are all SHPT patients who underwent PTX surgery in this center. Retrospectively collect the CT images of SHPT patients who completed preoperative scans with the same CT scanning equipment in our hospital from 2019 to 2023, and collect the general demographic information, biochemical indexes, surgery-related indexes and postoperative test indexes of SHPT patients.

[0063] The present invention intends to adopt a combined retrospective and prospective analysis method. To avoid the increase of statistical type I and type II errors caused by small sample sizes, and in combination with previous relevant retrospective CT radiomics studies, according to the inclusion criteria and exclusion criteria, 300 SHPT patients with complete preoperative CT images from 2019 to 2023 were retrospectively collected, the CT axial images of the above patients were collected, and a case database was established. 240 cases were randomly selected from this case database as the training set, and the remaining 60 cases were used as the validation set. Prospectively collect 100 SHPT patients who meet the inclusion and exclusion criteria in 2024 as the test set to evaluate the accuracy of the prognostic assessment model.

[0064] Specifically, the inclusion criteria (all of the following criteria need to be met):

[0065] * Requiring PTX treatment due to renal SHPT (serum i-PTH concentration continuously > 800 pg / mL, non-responsive to drug treatment, with severe osteoporosis and skeletal deformities;

[0066] *Preoperative ultrasound examination found that more than one parathyroid gland was enlarged, with a diameter ≥ 1 cm and rich blood supply;

[0067] *There were preoperative CT examination data;

[0068] *PTX surgery was successfully performed (defined as the number of resected parathyroid glands ≥ 3 and the serum i-PTH concentration < 60 pg / ml on the 1st - 3rd day after surgery).

[0069] Exclusion criteria (meeting any one of the following criteria):

[0070] *Combined with biliary, pancreatic or liver diseases;

[0071] *Using drugs such as Cinacalcet to treat renal SHPT within 6 months before surgery;

[0072] *Recurrent secondary PTX;

[0073] *Incomplete medical records;

[0074] *Combined with papillary thyroid carcinoma;

[0075] *Incomplete parathyroidectomy (the number of resected parathyroid glands is less than 3) or unsuccessful PTX surgery (serum i-PTH ≥ 60 pg / ml on the 1st - 3rd day after surgery);

[0076] *Parathyroid malignant tumor found by postoperative histological examination.

[0077] S102: Extracting EPH highly relevant features based on the parathyroid image for postoperative EPH prediction to obtain a prediction result;

[0078] Among them, the construction process of the EPH highly relevant features is as follows:

[0079] S1: Obtaining the parathyroid image set, EPH grouping label, and blood calcium concentration label of the patient;

[0080] S2: The parathyroid image set is subjected to feature extraction based on the EPH grouping label and blood calcium concentration label to obtain primary features;

[0081] S3: Feature screening is performed on the primary features to obtain the EPH highly relevant features.

[0082] In one embodiment, the S2 is specifically:

[0083] S21: The parathyroid image set is subjected to feature extraction based on the EPH grouping label to obtain the first radiomics feature; the first radiomics feature is subjected to feature extraction based on the blood calcium concentration label pair to obtain the primary feature; the blood calcium concentration label is a binary classification label, and is divided into a first blood calcium concentration and a second blood calcium concentration through a preset blood calcium concentration threshold;

[0084] In one embodiment, S21 is replaced with S22: The parathyroid image set is subjected to feature extraction based on the EPH grouping label to obtain the first radiomics feature; at the same time, the parathyroid image set is subjected to feature extraction based on the blood calcium concentration label to obtain the second radiomics feature; the intersection operation is performed on the first radiomics feature and the second radiomics feature to obtain the primary feature;

[0085] In one embodiment, S21 is replaced with S23: The parathyroid image set is subjected to feature extraction based on the blood calcium concentration label to obtain the second radiomics feature; the second radiomics feature is subjected to feature extraction based on the EPH grouping label to obtain the primary feature.

[0086] In one embodiment, the specific process of S2 is as follows:

[0087] S24: The parathyroid image set is subjected to feature extraction based on the EPH grouping label to obtain the first radiomics feature; the correlation calculation is performed between the first radiomics feature and the blood calcium concentration label to obtain the primary feature; the blood calcium concentration label is a continuous variable label;

[0088] In one embodiment, S24 is replaced with S25: The parathyroid image set is subjected to feature extraction based on the EPH grouping label to obtain the first radiomics feature; at the same time, the parathyroid image set is subjected to feature extraction based on the blood calcium concentration label to obtain the third radiomics feature; the correlation calculation is performed on the first radiomics feature and the third radiomics feature to obtain the primary feature;

[0089] In one embodiment, S24 is replaced with S26: The parathyroid image set is subjected to feature extraction based on the blood calcium concentration label to obtain the third radiomics feature, and the correlation calculation is performed between the third radiomics feature and the EPH grouping label to obtain the primary feature.

[0090] In one embodiment, the SHAP value is a method for explaining the predictions of a machine learning model, which is based on the Shapley value in game theory. In a machine learning model, each feature is regarded as a "player", and the Shapley value of the feature (i.e., the SHAP value) represents the average contribution of the feature in all possible combinations of features. The SHAP value is calculated by comparing the predictions of the model when a specific feature is present and absent, and this process is iterated for each feature and each sample in the dataset. The SHAP value assigns an importance value to each feature, providing a local and consistent explanation of the model's behavior, revealing which features have the greatest impact on a specific prediction, whether positive or negative. This is valuable for understanding the reasoning process behind complex machine learning models such as deep neural networks.

[0091] In one embodiment, the steps of S3 are as follows:

[0092] First step: Input the primary features into the machine learning model to perform preliminary sorting and screening on the features to obtain candidate features;

[0093] Second step: Analyze the contribution degree of the candidate features through SHAP values to obtain contribution degree features;

[0094] Third step: Screen out irrelevant features from the contribution degree features and sort them to obtain EPH highly correlated features.

[0095] In one embodiment, the machine learning model is one or several of the following: decision tree, GBDT, LightGBM, XGBoost, CatBoost.

[0096] In one embodiment, the process of extracting parathyroid imaging features to obtain parathyroid imaging omics features includes segmentation and extraction. The parathyroid imaging is segmented to obtain the parathyroid lesion area, and then the lesion size, shape, texture, edge, and functional information of the lesion area are extracted to obtain parathyroid imaging omics features.

[0097] In one embodiment, the method further includes stability evaluation. The imaging omics features are screened through stability evaluation to obtain features with good stability; based on the blood calcium concentration, the features with good stability are screened to obtain EPH highly correlated features. The stability evaluation judges the stability of the imaging omics features through the intraclass correlation coefficient and the coefficient of variation. When the intraclass correlation coefficient is greater than or equal to the preset intraclass threshold and the coefficient of variation is less than or equal to the preset variation threshold, it is determined that the stability is good and the features are retained; otherwise, they are screened out.

[0098] In one embodiment, the highly correlated features include one or more of the following: 3D m2 local binary pattern gray level size matrix gray non-uniformity normalization, 3D k local binary pattern gray level size matrix region entropy, wavelet LLL first-order skewness, 3D k local binary pattern gray level run matrix run length non-uniformity normalization, wavelet HLL first-order skewness.

[0099] In one embodiment, LBP (Local Binary Patterns) is a very simple but highly efficient local texture feature description operator; lbp_3Dm2_glszm GrayLevelNonUniformityNormalized (3Dm2 local binary pattern gray level size matrix gray non-uniformity normalization), lbp_3Dk_glszm_ZoneEntropy (3D k local binary pattern gray level size matrix region entropy), wavelet LLL first-order Skewness, lbp 3D k glrm RunLengthNonUniformityNormalized (3D k local binary pattern gray level run matrix run length non-uniformity normalization), wavelet HLL first-order Skewness.

[0100] In one embodiment, the method further includes obtaining clinical features. First, based on the EPH grouping labels, the data is divided into EPH group and non-EPH group and univariate analysis is performed to obtain statistically significant features. Then, multivariate analysis is performed on the statistically significant features of the two groups to obtain clinical features. The clinical features are fused with the EPH highly correlated features and then postoperative EPH prediction is performed; the clinical features include one or more of the following: parathyroid calcification, hemoglobin, alkaline phosphatase, preoperative parathyroid hormone.

[0101] In one embodiment, the method further includes multi-modal data feature prediction. First, the patient's clinical information and biochemical indicators are obtained and feature extraction is performed to obtain clinical information features and biochemical features. Based on the clinical information features, biochemical features and the EPH highly correlated features, postoperative EPH prediction is performed to obtain a prediction result.

[0102] In a specific embodiment, image segmentation is performed on the CT images of 240 patients, and radiomics features that are highly correlated with prognosis and have reliable stability are extracted and screened. The present invention will use the selected radiomics features to construct a radiomics model, and combine clinical information and radiomics features to construct a comprehensive evaluation model for predicting the risk of EPH occurrence in SHPT patients. The data of the remaining 60 patients is used to verify and optimize the two models.

[0103] Extraction and screening of radiomics features: The lesion regions of the CT images of SHPT patients were outlined using image segmentation software. The outlining was performed independently by two clinical doctors to ensure the repeatability of the radiomics features. Python software tools were used to extract radiomics features from the outlined regions, verify the stability and reliability of the radiomics features, and then further screen the extracted massive radiomics features through different machine learning methods (such as mRMR feature selection algorithm, Lasso-Logist regression model, etc.), so as to preliminarily determine the imaging features that can be used as evaluation indicators for predicting the occurrence of EPH after SHPT surgery.

[0104] Specifically, extraction of radiomics features of parathyroid lesions in SHPT patients: The CT images in the PACS system were exported in DICOM format and imported into 3D Slicer software. Doctors in our department manually outlined along the lesion edges to obtain the volume of interest (VOI). The segmented images are as Figure 4 shown. Then, with the help of Python 3.6.5 software tools, quantitative features including the size, shape, texture, edge, and function of the lesion were extracted from the VOI in a high-throughput manner. Among them, the texture included multi-dimensional decomposition of the lesion image, and entropy matrices, co-occurrence matrices, run-length matrices, etc. were obtained through wavelet decomposition, Gabor filtering, etc. to describe the ruggedness and chaos of the tumor surface texture, and then converted into data that can be collected and analyzed.

[0105] Furthermore, in a specific embodiment, evaluation of the stability of radiomics features: The stability and reliability of radiomics features are the prerequisite for clinical applications. The present invention intends to adopt the following method to evaluate the stability and reliability of radiomics features of the training set CT images and screen out features with good robustness. From 240 patients in the training set, 30 patients were randomly selected, and a doctor segmented the lesions twice. Another doctor segmented these 30 patients again. Features were extracted after segmentation. For the extraction results of the two segmentations by the same doctor, as well as the extraction results of the two doctors' separate segmentations, the intraclass correlation coefficient (ICC) and coefficient of variation (CV) were used to evaluate the feature stability. ICC≥0.8 and CV≤10% indicate good stability.

[0106] Further, in a specific embodiment, screening of key radiomics features for the risk of EPH after PTX in SHPT patients: There is still a large amount of redundant information among the screened features with reliable correctability and stability. It is necessary to further screen out the key radiomics features that have a good correlation with predicting the risk of EPH after SHPT. The specific scheme is as follows:

[0107] (1) In this study, serum calcium levels will be monitored continuously for 4 days after surgery. If the serum calcium level in any one monitoring is lower than 2.1 mmol / L, it is the EPH group; otherwise, it is the non-EPH group.

[0108] (2) First, the Lasso-Logist regression model is used to compress the coefficients of some features to 0 through variable coefficient compression, so as to screen out these features. Then, the maximum relevance and minimum redundancy (mRMR) of the mRMRe software package based on R language is used to rank the remaining features, and the top 10% of the features are selected as candidate features.

[0109] (3) Handling class sample imbalance: In the training data, the sample sizes of SHPT with different prognoses may vary, which will affect the prediction accuracy of the model. And the strategy of simply replicating samples to increase the minority class samples will lead to overfitting of the model and affect the generalization of the model. The basic idea of the SMOTE (Synthetic Minority over Sampling Technique) algorithm is to analyze the minority class samples and artificially synthesize new samples according to the minority class samples and add them to the dataset, which is an effective method to solve unbalanced samples. In this study, we plan to use DMwR based on R language to perform SMOTE analysis to process the data and reduce the imbalance of class samples.

[0110] (4) Import the candidate features processed by SMOTE into the R language software. Then, the Lasso-Logist regression model of the glmnet software package based on R language is used to further screen the features, and finally, the radiomics features highly correlated with the occurrence of EPH after surgery in SHPT patients are selected.

[0111] In another specific embodiment, in the traditional radiomics feature extraction process, first, univariate analysis is performed on all features according to the prediction results to identify features with significant statistical differences. Subsequently, correlation analysis is carried out among the features to eliminate highly correlated features to avoid interference of the collinearity problem on model building. Further, the Lasso method is applied to screen the features. Lasso is a regularization method that can effectively reduce the weights of unimportant features and helps improve the generalization ability of the model. Through these steps, the most predictive and informative feature set is screened out, which plays a key role in building machine learning or deep learning models. After selecting the finally retained features, they are used to construct a traditional radiomics model (Trad rad model).

[0112] In another specific embodiment, the present invention innovatively proposes a radiomics feature extraction scheme. First, the LightGBM model is used to preliminarily rank and screen the features to identify features that have an important impact on the target variable. Subsequently, the contribution degree and influence degree of each feature to the model prediction are analyzed through SHAP (SHapley Additive exPlanations) values, so as to better understand the relationships and importance among the features. Based on the analysis results of the SHAP values, the top 20 most important features are selected and retained. By further applying the LASSO method to screen these 20 important features, the finally determined set of key features is retained for constructing a new method radiomics model (New rad model).

[0113] The entire scheme combines a machine learning model (LightGBM), SHAP value analysis, and LASSO screening method. This method helps improve model performance, reduce the risk of overfitting, and provides a powerful tool for deeply understanding the laws behind medical image data.

[0114] In a specific embodiment, a total of 1834 radiomics features are extracted from parathyroid images. For the traditional radiomics feature processing method, univariate screening is performed, and features with p < 0.001 are retained. A total of 41 features are screened out. Using the correlation coefficient method, 26 remaining features are obtained. Finally, 6 radiomics features are retained through LASSO screening, as Figure 5 shown. During the feature screening process of the new method, the LightGBM is used to screen and retain the top 20 important features, as Figure 6 shown. These 20 features are further incorporated into LASSO for screening, and finally 5 important features are retained, as Figure 7 shown.

[0115] In one embodiment, LightGBM (Light Gradient Boosting Machine) is a gradient boosting framework that uses decision trees as base learners. LightGBM is designed for efficient parallel computing, and its "Light" is reflected in the following aspects: faster training speed, lower memory usage, support for single-machine multi-threading and multi-machine parallel computing, and the ability to handle large-scale data.

[0116] In one embodiment, the prediction is performed by one or more of the following methods: multi-layer feedforward neural network, partial least squares regression method, random forest, support vector machine, Boost, and Logistic regression model.

[0117] In a specific embodiment, the construction, validation, and testing of the radiomics prognosis assessment model and the comprehensive assessment model: According to the occurrence of EPH in all enrolled patients, further screen the CT radiomics features highly correlated with EPH. Use various mathematical methods (multi-layer feedforward neural network method, partial least squares regression method, random forest method, support vector machine, Boost algorithm, Logist regression model) to judge the degree of association between the radiomics features and the occurrence of EPH, and select the prediction model as the best model for predicting EPH after PTX in SHPT patients through optimization, validation, and testing.

[0118] Specifically, the construction of the radiomics model and the comprehensive assessment model: For the radiomics features highly correlated with the occurrence of EPH in SHPT patients selected by the present invention, use various mathematical methods based on the rms software package in R language to judge their degree of association with EPH:

[0119] 1) Multi-layer feedforward neural network (back-propagation, BP) method: It is the most widely used and mature neural network model in current clinical research.

[0120] 2) Partial least squares regression method (partial least squares regression, PLSR): A new type of multivariate statistical data analysis method that mainly studies the regression modeling of multiple dependent variables on multiple independent variables. It is more effective to use PLSR especially when there is a high degree of linear correlation within each variable.

[0121] 3) Random Forest method (Random Forest, RF): This algorithm contains multiple decision trees, can have thousands of input variables, and can also rank the importance of variables for the classification result.

[0122] 4) Support Vector Machine (SVM): Its foundation is statistical learning theory. It adopts the principle of minimizing structural risk, seeking the best compromise between the complexity of the model and the learning ability based on limited sample information, and obtaining good generalization ability;

[0123] 5) Boost algorithm: Its main principle is to change the model of data distribution and combine weak classifiers obtained from different training sets into a strong classifier. Its basic idea is to assign corresponding weights to each variable according to the previous training results (including the overall classification accuracy). The advantage of this method is that it can focus on those samples that are difficult to separate and is not prone to overfitting;

[0124] 6) Logistic regression model (abbreviated as Logist model): It is a generalized linear regression analysis model, commonly used in data mining, automatic disease diagnosis and other fields. For example, exploring the risk factors causing diseases and predicting the probability of disease occurrence based on risk factors, etc.

[0125] Finally, based on the correlation results between EPH and radiomics features, using ROC (receiver operating characteristic) curve analysis, taking the area under the curve (AUC), prediction accuracy, etc. as evaluation indicators of the prediction model, comparing the performance differences of the above six prediction models, and selecting the prediction model with a large AUC value and high prediction accuracy as the best model for predicting and evaluating the occurrence of EPH after SHPT surgery.

[0126] In a specific embodiment, a clinical model (Clin model) is established: According to whether postoperative hypocalcemia occurs, the patients in the training set are divided into two groups, and the relevant clinical indicators are analyzed by univariate analysis. The features with statistical significance (p < 0.05) are retained, and further multivariate analysis is carried out. The clinical features with significance in multivariate analysis are retained for clinical model training (p < 0.05). After univariate analysis, the features of multivariate analysis are retained, and the results are shown in Table 1.

[0127] Table 1 Results of multivariate analysis of clinical features

[0128]

[0129] In a specific embodiment, the fusion model is established as follows: The trad rad model and the new rad model are respectively fused with the clin model, and the model performance is verified through an external test set. The area under the ROC curve (AUC) is calculated, and the effectiveness of this feature screening method is examined by comparing the AUCs. This research design that integrates multiple data sources, multiple feature selection methods, and conducts comparative evaluations helps to verify the effectiveness of the feature screening method and provides important guidance for future medical imaging omics research.

[0130] In a specific embodiment, prediction models are trained with features screened by different methods. The performance of the models is shown in Table 2. The features screened by different methods include: clinical features, features screened by traditional omics methods, features screened by the new method, clinical features + features screened by traditional omics methods, and clinical features + features screened by the new method. Each feature is divided into a training set and a test set before training the prediction model. After training is completed using the training set, the test set is then used to test the prediction model. The model results of the training set and the test set are respectively as Figure 8 、 Figure 9 shown.

[0131] Table 2 Performance of prediction models trained with different features

[0132]

[0133] In a specific embodiment, the present invention also comprehensively incorporates the imaging omics features, clinical, and blood biochemical indicators of patients for analysis. The rank sum (log-rank) hypothesis test is performed on ranked data and continuous data. For imaging omics features with significant statistical differences, the receiver operating characteristic (ROC) curve is plotted, the area under the curve (AUC) is judged, and variables meaningful for the prognosis assessment of SHPT are screened out. Then, multiple mathematical methods are used for analysis and attempts are made to build models. Finally, the optimal comprehensive evaluation model with the best prediction performance is selected.

[0134] In a specific embodiment, the verification and optimization of the imaging omics model and the comprehensive evaluation model: This project plans to use 60 retrospectively collected SHPT patients as the validation set, collect preoperative CT data, clinical, blood biochemical, and postoperative information, and verify the prediction performance of the imaging omics model and the comprehensive evaluation model that incorporates clinical information and biochemical indicators. The concordance index (C-index) is calculated using the Hmisc software package based on R language to evaluate the evaluation performance of the two models for predicting EPH after SHPT surgery. The calibration curve is plotted using the rms software package based on R language to evaluate the fitting degree of the two models for the training data and the validation data. According to the verification results and the calibration curve results, the imaging omics model and the comprehensive evaluation model are further adjusted and optimized.

[0135] In a specific embodiment, testing of the radiomics model and the comprehensive evaluation model: 100 SHPT patients who underwent imaging examinations and surgeries in our hospital were prospectively and continuously collected. Preoperative CT images, clinical, biochemical, and postoperative information of the patients were collected and sorted out to test the accuracy of the radiomics prognosis evaluation model for SHPT patients and the comprehensive evaluation model incorporating clinical information and biochemical indicators. The inclusion criteria for prospectively collected cases were the same as those for retrospectively collected cases. The variables finally included in the two models were imported into R language, and nomograms were constructed using the rms software package based on R language, with the aim of providing a more intuitive and reliable tool for predicting the risk of EPH in SHPT patients.

[0136] In a specific embodiment, the construction, verification, and testing of the radiomics model and the comprehensive evaluation model were carried out using Python software. Modeling was performed through six mathematical algorithms: multi-layer feedforward neural network method, partial least squares regression method, random forest method, support vector machine, Boost algorithm, and Logist regression model. The model with the best prediction performance was selected and verified using the case data of the validation set. The diagnostic efficacy of the six prediction models was evaluated and compared using the ROC curve. The maximum efficacy of the quantitative parameters was analyzed through the area under the curve AUC. The time when the Youden index was the largest was selected as the threshold, and its diagnostic accuracy, sensitivity, and specificity were calculated.

[0137] In a specific embodiment, all statistical analyses were processed using SPSS software (IBM SPSS Statistics for Windows, Version 22.0. Armonk, NY: IBM Corp) and R language software (R Studio, Version 1.1.456; R, Version 3.5.1). A P value < 0.05 was considered statistically significant. Normality tests and homogeneity of variance tests were performed on measurement data. Measurement data were expressed as mean ± standard deviation (following a normal distribution) or median and interquartile range.

[0138] Figure 2 A schematic diagram of an image-based EPH prediction system after parathyroidectomy provided by an embodiment of the present invention specifically includes:

[0139] Acquisition unit: acquiring parathyroid images of a patient;

[0140] Prediction unit: extracting EPH highly correlated features based on the parathyroid images to predict postoperative EPH and obtaining a prediction result;

[0141] Among them, the construction process of the EPH highly correlated features is as follows:

[0142] S1: acquiring parathyroid image sets and blood calcium concentration data sets of EPH patients and non-EPH patients;

[0143] S2: Extract primary features of the EPH group and the non-EPH group based on the parathyroid image set and the blood calcium concentration data set;

[0144] S3: Screen the primary features of the EPH group and the non-EPH group to obtain the EPH highly correlated features.

[0145] Figure 3 The schematic diagram of an EPH prediction device after parathyroidectomy based on images provided by the embodiments of the present invention specifically includes:

[0146] A memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, any one of the above-mentioned methods for predicting EPH after parathyroidectomy based on images.

[0147] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, any one of the above-mentioned methods for predicting EPH after parathyroidectomy based on images.

[0148] The verification results of this verification embodiment show that assigning fixed weights to the indications can improve the performance of this method compared to the default settings. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein. In several embodiments provided by the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways.

[0149] For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling, direct coupling, or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in electrical, mechanical, or other forms. The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in each embodiment of the present invention, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit. Those of ordinary skill in the art can understand that all or part of the steps in the above-described methods of the embodiments can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium, and the storage medium can include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk, or optical disc, etc.

[0150] Those of ordinary skill in the art can understand that all or part of the steps in the above-described methods of the embodiments can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium, and the above-mentioned medium storage can be read-only memory, magnetic disk, or optical disc, etc.

[0151] The above provides a detailed introduction to a computer device provided by the present invention. For those of ordinary skill in the art, according to the idea of the embodiments of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An image-based method for predicting EPH after parathyroidectomy, characterized in that: The method comprises: Obtain images of the patient's parathyroid glands; Extracting EPH highly correlated features based on the parathyroid image; the method also includes obtaining clinical features, first dividing the data into EPH group and non-EPH group based on the EPH grouping label and performing univariate analysis to obtain statistically significant features, then performing multivariate analysis on the two groups of statistically significant features to obtain clinical features, and fusing the clinical features with the EPH highly correlated features to predict postoperative EPH; The construction process of the EPH highly correlated feature is as follows: S1: Obtain the patient's parathyroid imaging set, EPH grouping label, and blood calcium concentration label; S2: The parathyroid image set is feature extracted based on the EPH grouping label and the blood calcium concentration label to obtain primary features; S3: Screening the primary features to obtain the EPH highly correlated features; The highly correlated features are: 3D m2 local binary pattern grayscale level size matrix grayscale non-uniformity normalization, 3Dk local binary pattern grayscale level size matrix regional entropy, wavelet LLL first-order skewness, 3D k local binary pattern grayscale level run matrix run length non-uniformity normalization, wavelet HLL first-order skewness; The clinical features are: parathyroid calcification, hemoglobin, alkaline phosphatase, and preoperative parathyroid hormone.

2. The method for predicting EPH after parathyroidectomy based on imaging according to claim 1, characterized in that: The S2 is specifically: S21: extracting features from the parathyroid image set based on the EPH grouping label to obtain a first radiomics feature; extracting features from the first radiomics feature based on the blood calcium concentration label to obtain a primary feature; The blood calcium concentration label is a binary label, which is divided into a first blood calcium concentration and a second blood calcium concentration by a preset blood calcium concentration threshold.

3. The method for predicting EPH after parathyroidectomy based on imaging according to claim 2, characterized in that: The S21 is replaced by S22: the parathyroid image set is feature extracted based on the EPH grouping label to obtain a first imaging feature; at the same time, the parathyroid image set is feature extracted based on the blood calcium concentration label to obtain a second imaging feature; the first imaging feature and the second imaging feature are intersected to obtain a primary feature.

4. The method for predicting EPH after parathyroidectomy based on imaging according to claim 2, characterized in that: The S21 is replaced by S23: the parathyroid image set is feature extracted based on the blood calcium concentration label to obtain a second radiomics feature; the second radiomics feature is feature extracted based on the EPH grouping label to obtain a primary feature.

5. The method for predicting EPH after parathyroidectomy based on imaging according to claim 1, characterized in that: The specific process of S2 is: S24: extracting features from the parathyroid image set based on the EPH grouping label to obtain a first radiomics feature; performing correlation calculation between the first radiomics feature and the blood calcium concentration label to obtain a primary feature; The blood calcium concentration label is a continuous variable label.

6. The method for predicting EPH after parathyroidectomy based on imaging according to claim 5, characterized in that: The S24 is replaced by S25: the parathyroid image set is feature extracted based on the EPH grouping label to obtain a first imaging omics feature; at the same time, the parathyroid image set is feature extracted based on the blood calcium concentration label to obtain a third imaging omics feature; the first imaging omics feature and the third imaging omics feature are correlated to obtain a primary feature.

7. The method for predicting EPH after parathyroidectomy based on imaging according to claim 5, characterized in that: The S24 is replaced by S26: the parathyroid image set is feature extracted based on the blood calcium concentration label to obtain a third radiomics feature, and the third radiomics feature is correlated with the EPH grouping label to obtain a primary feature.

8. The method for predicting EPH after parathyroidectomy based on imaging according to any one of claims 1 to 7, characterized in that: The steps of S3 are: Step 1: Input the primary features into the machine learning model to preliminarily sort and screen the features to obtain candidate features; Step 2: Analyze the contribution of the candidate features through the SHAP value to obtain the contribution feature; Step 3: Eliminate irrelevant features from the contribution features and sort them to obtain EPH highly relevant features.

9. An image-based prediction system for EPH after parathyroidectomy, characterized in that: include: When the system is executed, the image-based method for predicting EPH after parathyroidectomy according to any one of claims 1 to 8 is implemented.

10. An imaging-based device for predicting EPH after parathyroidectomy, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, the image-based method for predicting EPH after parathyroidectomy according to any one of claims 1 to 8 is implemented.

11. A computer-readable storage medium having a computer program thereon, characterized in that: include: When the computer program is executed by a processor, the image-based method for predicting EPH after parathyroidectomy according to any one of claims 1 to 8 is implemented.

Images