Clinical chemotherapy effect prediction analysis method based on multi-center MRI image interpretable model

Abstract

The invention discloses a clinical chemotherapy curative effect prediction and analysis method based on a multi-center MRI image interpretable model, and relates to the technical field of medical image analysis, and the method is characterized in that an MRI image of a patient is collected, the MRI image is screened and preprocessed, radiomics characteristics are extracted, and a clinical chemotherapy curative effect prediction and analysis result is obtained. On the basis of an intra-group correlation coefficient ICC, Mann-Whitney U test, Spearman correlation analysis and an elastic network regression algorithm, nasopharynx cancer radiomics characteristics are screened out from the radiomics characteristics to construct a GNN neural network model so as to predict the induced chemotherapy reaction of the patient with the local advanced nasopharynx cancer. The core problem that the curative effect of a patient with local advanced nasopharyngeal carcinoma cannot be accurately predicted before induced chemotherapy at present is solved, and chemotherapy injury and excessive treatment caused by induced chemotherapy are effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and more specifically, to a clinical chemotherapy efficacy prediction and analysis method based on a multi-center MRI image interpretable model. Background Art

[0002] Nasopharyngeal carcinoma is an epithelial cancer originating from the nasopharyngeal mucosal lining. Its global geographical distribution is extremely uneven, mainly concentrated in East Asia and Southeast Asia. Early nasopharyngeal carcinoma is treated with radiotherapy as the main treatment method. Induction chemotherapy has played a superior role in reducing tumor burden and improving survival in the treatment of locally advanced nasopharyngeal carcinoma. Although high-risk nasopharyngeal carcinoma patients with distant metastasis may benefit from additional induction chemotherapy, clinical trials have shown that different patients with locally advanced nasopharyngeal carcinoma have large differences in their response to induction chemotherapy. About 9% of patients with locally advanced nasopharyngeal carcinoma show poor tumor response to induction chemotherapy, about 23% of patients with locally advanced nasopharyngeal carcinoma cannot benefit from induction chemotherapy, and about 20% to 30% of patients with locally advanced nasopharyngeal carcinoma will have local recurrence or distant metastasis. Therefore, accurate preoperative prediction of the response of patients with locally advanced nasopharyngeal carcinoma to induction chemotherapy will not only help to develop personalized treatment plans for patients with locally advanced nasopharyngeal carcinoma and avoid delaying the best treatment time; it can also prevent low-risk nasopharyngeal carcinoma patients from receiving excessive treatment, thereby reducing their economic burden and toxic side effects. Summary of the invention

[0003] In view of the current problems that the efficacy of induction chemotherapy cannot be accurately predicted before induction chemotherapy for patients with locally advanced nasopharyngeal carcinoma, and the possible delay in the optimal treatment time and overtreatment, the purpose of the present invention is to provide a clinical chemotherapy efficacy prediction and analysis method based on a multi-center MRI image interpretable model to predict the induction chemotherapy response of patients with locally advanced nasopharyngeal carcinoma, which will help enhance the confidence of clinicians in treatment decisions.

[0004] To achieve the above object, the present invention provides the following technical solution: a clinical chemotherapy efficacy prediction and analysis method based on a multi-center MRI image interpretable model, which comprises the following steps:

[0005] S1. Patients with nasopharyngeal carcinoma diagnosed by pathology in multiple hospitals are selected as analysis subjects, and data of the analysis subjects are obtained, the data including MRI images before induction chemotherapy and at n weeks of induction chemotherapy, where n is a given value. The MRI images include T2WI and CET1WI.

[0006] S2. Remove unqualified data, and use the SimpleITK package to process the MRI images in the remaining data to obtain qualified data. The processing includes N4 bias field correction and standardized voxel spacing. The removal criteria include: 1. For the nasopharyngeal carcinoma primary tumor that cannot be measured in the MRI image before induction chemotherapy but is determined based on pathological analysis, remove the MRI image of the location of the primary nasopharyngeal carcinoma tumor. 2. Remove MRI images that cannot be analyzed by imaging genomics. 3. Remove the data of the analysis object without MRI images before induction chemotherapy. 4. Remove the data of the analysis object who did not receive complete induction chemotherapy. 5. Remove the data of the analysis object with a history of anti-tumor treatment before the first MRI examination. 6. Remove the data of the analysis object with existing or concurrent other primary malignant tumors.

[0007] S3. Randomly select at least 50 MRI images before induction chemotherapy from the qualified data as stability assessment data. Select multiple radiologists, and each radiologist independently segmented the stability assessment data to segment the region of interest of nasopharyngeal carcinoma tumors. Use the Pyradiomics package to extract the radiomic feature data of the region of interest segmented by each radiologist. The radiomic feature data includes the type and the eigenvalue corresponding to each type, and the intraclass correlation coefficient ICC is calculated for each extracted radiomic feature type. Given a stability threshold A, the radiomic feature type with an ICC greater than A is regarded as a stable radiomic feature type.

[0008] S4. Multiple radiologists discussed and processed the MRI images before induction chemotherapy in the qualified materials to segment the regions of interest of nasopharyngeal carcinoma tumors. The Pyradiomics package was used to extract feature values ​​of stable radiomics feature types from the regions of interest segmented by multiple radiologists.

[0009] S5. Given a threshold value of B% for efficacy. According to qualified data, for a patient, if the volume of the primary nasopharyngeal carcinoma tumor in the MRI image at the time of induction chemotherapy for n weeks is reduced by more than B% compared with the volume of the primary nasopharyngeal carcinoma tumor in the MRI image before induction chemotherapy, then the induction chemotherapy efficacy value of the patient is C. Otherwise, the induction chemotherapy efficacy value of the patient is D. Both C and D are given values, C>D.

[0010] S6. Mann-Whitney U test and Spearman correlation analysis were performed according to the characteristic values ​​of stable radiomics feature types to screen out stable radiomics feature types with a P value less than 0.05 and a Spearman correlation coefficient greater than 0.6 in the Mann-Whitney U test results.

[0011] S7. Use the glmnet function and cv.glmnet function to screen the stable radiomics feature types screened out in S6 based on the elastic network regression algorithm. The formula of the glmnet function is as follows:

[0012]

[0013] Where α is 1, λ is determined by the cv.glmnet function, w is the regression coefficient for the stable radiomics feature type, and x i w is the characteristic value of the i-th patient corresponding to the stable radiomics feature type, y i For x i The corresponding induction chemotherapy efficacy value of the patient, N is the total number of analysis subjects in the qualified data.

[0014] Calculate w for each stable radiomics feature type screened by S6. If a stable radiomics feature type does not have w, remove the stable radiomics feature type. If a stable radiomics feature type has w, retain the stable radiomics feature type as the radiomics feature type of nasopharyngeal carcinoma.

[0015] S8. Using the characteristic value of the radiomics feature type of nasopharyngeal carcinoma as the independent variable and the efficacy of induction chemotherapy for patients determined by S5 as the dependent variable, a GNN neural network model was trained to predict the efficacy of induction chemotherapy for patients with nasopharyngeal carcinoma.

[0016] The present invention is further configured as follows: A≥0.75 in S3.

[0017] The present invention is further configured as follows: n≥2 in S1, and B%≥30% in S5.

[0018] The present invention is further configured as follows: C is 1 and D is 0 in S5.

[0019] The present invention is further configured as follows: the radiomics feature types of nasopharyngeal carcinoma include kurtosis, cluster prominence, grayscale non-uniformity, cluster trend, dependency entropy, maximum correlation coefficient, contrast and intensity of T2WI, and minimum value, skewness, cluster shadow, grayscale non-uniformity, regional entropy and kurtosis of CET1WI.

[0020] The present invention is further configured as follows: S5 also includes copying the analysis object data with the induction chemotherapy efficacy value D by integer multiples, so that the ratio between the number of analysis object data with the induction chemotherapy efficacy value D and the number of analysis object data with the induction chemotherapy efficacy value C is closest to 1:1.

[0021] The present invention is further configured as follows: S4 also includes performing batch correction on the extracted characteristic values ​​to remove batch effects caused by different hospitals and different scanners.

[0022] The present invention is further configured such that: in S4, batch correction adopts at least one of a combat algorithm and a z-score method.

[0023] The present invention is further configured as follows: the data in S1 also include clinical pathological data before induction chemotherapy and at n weeks after induction chemotherapy, and the clinical pathological factors in the clinical pathological data include age, gender, smoking history, family medical history, weight, height, body mass index, hemoglobin concentration, white blood cell concentration, neutrophil concentration, lymphocyte concentration, neutrophil-lymphocyte ratio, platelet concentration, albumin concentration, EBV-DNA concentration, clinical T stage, clinical N stage and WHO type.

[0024] The removal criteria for S2 also include 7. Removal of analysis subjects without clinical pathological data before induction chemotherapy.

[0025] The following steps are also included:

[0026] S9. Using the clinical pathological factors in S1 as independent variables and the efficacy of induction chemotherapy for patients determined by S5 as the dependent variable, univariate and multivariate regression analysis was performed to determine multiple clinical pathological factors with the greatest correlation with the efficacy of induction chemotherapy for patients as clinical pathological factors for nasopharyngeal carcinoma.

[0027] S10. Using the prediction results of the GNN neural network model in S8 and the data of clinical pathological factors of nasopharyngeal carcinoma in the clinical pathological data of S1 as independent variables, and the efficacy of induction chemotherapy for patients determined in S5 as the dependent variable, a machine learning model for predicting the efficacy of induction chemotherapy for nasopharyngeal carcinoma patients is trained, wherein the machine learning model is at least one of a GNN neural network model, an LR model, and a DT model.

[0028] The present invention is further configured as follows: the clinical pathological factors of nasopharyngeal carcinoma include hemoglobin concentration, neutrophil concentration and EBV-DNA concentration.

[0029] In summary, the present invention has the following beneficial effects compared with the prior art: the present invention provides a clinical chemotherapy efficacy prediction and analysis method based on a multi-center MRI image interpretable model, by collecting MRI images of patients, screening and preprocessing the MRI images, extracting radiomic features, and based on the intraclass correlation coefficient ICC, Mann-Whitney U test, Spearman correlation analysis and elastic network regression algorithm, screening nasopharyngeal carcinoma radiomic features from radiomic features to construct a GNN neural network model to predict the induction chemotherapy response of patients with locally advanced nasopharyngeal carcinoma, which will help enhance the confidence of clinicians in treatment decisions, help clinicians determine the best time for treatment, and reduce the occurrence of over-medicalization. At the same time, the present invention also establishes a machine learning model based on the prediction results of the GNN neural network model combined with clinical pathological factors, which further improves the prediction accuracy of the induction chemotherapy response of patients with locally advanced nasopharyngeal carcinoma. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the overall process of the embodiment.

[0031] Figure 2 A schematic diagram of the process of constructing a machine learning model in an embodiment.

[0032] Figure 3 The comparison results in (1) are explained by the model;

[0033] Figure 4 The comparison results in (2) are explained by the model;

[0034] Figure 5 The Kaplan-Meier curves for the training cohort in the model interpretability analysis (3) are shown;

[0035] Figure 6 Kaplan-Meier curves for the internal validation cohort in the model interpretability analysis (3);

[0036] Figure 7 Kaplan-Meier curves for the external validation cohort in the model interpretability analysis (3);

[0037] Figure 8 Kaplan-Meier curves for the cross-field strength validation cohort in the model interpretability analysis (3);

[0038] Fig. 9 The comparison results in (4) are for model interpretability analysis;

[0039] Fig.10 This is the SHAP analysis diagram of the top 10 contributing nasopharyngeal carcinoma radiomics feature types in the GNN neural network model of S8; DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention, and all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the invention.

[0041] It should be noted that the directions or positional relationships indicated by terms such as “center”, “up”, “down”, “horizontal”, “left”, “right”, “front”, “back”, “lateral” and “longitudinal” are based on the directions or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as a limitation on the present invention.

[0042] The term "region of interest" refers to the area of ​​the MRI image that is associated with the tumor. The term "T2WI" refers to T2-weighted imaging. The term "CET1WI" refers to contrast-enhanced T1-weighted imaging.

[0043] Example

[0044] like Figure 1 and Figure 2 As shown, this embodiment provides a clinical chemotherapy efficacy prediction and analysis method based on a multi-center MRI image interpretable model, which includes the following steps:

[0045] S1. Patients with nasopharyngeal carcinoma diagnosed by pathology in multiple hospitals were selected as analysis subjects, and the data of the analysis subjects were obtained, the data including MRI images before induction chemotherapy and at n weeks of induction chemotherapy, and clinical pathological data before induction chemotherapy and at n weeks of induction chemotherapy, where n is a given value. The MRI images include T2WI and CET1WI. The clinical pathological factors in the clinical pathological data include age, gender, smoking history, family history, weight, height, body mass index, hemoglobin concentration, white blood cell concentration, neutrophil concentration, lymphocyte concentration, neutrophil-lymphocyte ratio, platelet concentration, albumin concentration, EBV-DNA concentration, clinical T stage, clinical N stage and WHO type.

[0046] Specifically, n≥2, and n is 2 in this embodiment.

[0047] In this embodiment, 1077 analysis subjects were selected from three hospitals, and the MRI images in the data of these analysis subjects were obtained by a 3.0T scanner.

[0048] S2. Remove unqualified data, and use the SimpleITK package to process the MRI images in the remaining data to obtain qualified data, wherein the processing includes N4 bias field correction and standardized voxel spacing. N4 bias field correction is used to correct the inhomogeneity of magnetic field strength of different scanners. Standardized voxel spacing is used to improve the spatial consistency of MRI images. In this embodiment, the MRI image is resampled to 1×1×1mm3 by the linear interpolation algorithm in the SimpleITK package to standardize the voxel spacing. The SimpleITK package used in this embodiment is the SimpleITK package in the Python software.

[0049] The removal criteria include: 1. For the nasopharyngeal carcinoma primary tumor that cannot be measured in the MRI image before induction chemotherapy but is determined based on pathological analysis, the MRI image at the location of the nasopharyngeal carcinoma primary tumor is removed. In this embodiment, the nasopharyngeal carcinoma primary tumor with a diameter of less than 5 mm is identified as an unmeasurable nasopharyngeal carcinoma primary tumor. 2. MRI images that cannot be analyzed by imaging genomics are removed, that is, MRI images of poor quality are removed. 3. Analysis object data without MRI images before induction chemotherapy are removed. 4. Analysis object data that have not received complete induction chemotherapy are removed. 5. Analysis object data with a history of anti-tumor treatment before the first MRI examination are removed. 6. Analysis object data with existing or concurrent other primary malignant tumors are removed. 7. Analysis object data without clinical pathological data before induction chemotherapy are removed.

[0050] S3. Randomly extract at least 50 MRI images before induction chemotherapy from the qualified data as stability assessment data. In this embodiment, 50 MRI images before induction chemotherapy were randomly extracted from the qualified data as stability assessment data. Select multiple radiologists, each radiologist independently segmented the stability assessment data to segment the region of interest of the nasopharyngeal carcinoma tumor. In this embodiment, the segmentation process was performed using ITK-SNAP software. The Pyradiomics package was used to extract radiomics feature data from the region of interest segmented by each radiologist. The Pyradiomics package used in this embodiment is the Pyradiomics package in Python software. The radiomics feature data includes types and feature values ​​corresponding to each type. In this embodiment, 944 feature values ​​of radiomics feature types were extracted for T2WI and CET1WI, including 18 first-order statistical feature types, 14 shape-based feature types, 75 texture feature types, 93 square transform feature types, and 744 wavelet feature types. Then, the intraclass correlation coefficient ICC was calculated for each extracted radiomics feature type. Given a stability threshold A, the radiomics feature types with ICC greater than A are regarded as stable radiomics feature types.

[0051] Specifically, the ICC calculation formula is as follows:

[0052]

[0053] Wherein, k is the number of radiologists, and in this embodiment, k is 2; msb is the mean square between groups, which is equal to the variance of the eigenvalues ​​caused by different segmentation results of different radiologists on the same MRI image; mse is the mean square error, which is equal to the variance of the eigenvalues ​​caused by different segmentation results of the same reader on different MRI images.

[0054] Specifically, A≥0.75, and in this embodiment, A is 0.75.

[0055] In this embodiment, two radiologists (reader A with 8 years of experience in head and neck MRI interpretation and reader B with 25 years of experience in head and neck MRI interpretation) independently segmented the stability assessment data.

[0056] S4. Multiple radiologists discuss and segment the MRI images before induction chemotherapy in the qualified data to segment the region of interest of the nasopharyngeal carcinoma tumor. In this embodiment, the segmentation process is performed using ITK-SNAP software. The Pyradiomics package is used to extract feature values ​​for the stable radiomics feature types from the region of interest segmented by multiple radiologists.

[0057] In this embodiment, a radiologist (reader A with 8 years of experience in interpreting head and neck MRI) segments the region of interest of the nasopharyngeal carcinoma tumor, and another radiologist (reader B with 25 years of experience in interpreting head and neck MRI) confirms the segmented region of interest of the nasopharyngeal carcinoma tumor. If the two radiologists disagree on the segmentation results, the two radiologists discuss and use the Pyradiomics package to extract feature values ​​of stable radiomics feature types for the region of interest of the nasopharyngeal carcinoma tumor determined based on the consensus of the two radiologists.

[0058] This embodiment also performs batch correction on the extracted eigenvalues ​​to remove the batch effects caused by different hospitals and different scanners. The batch correction uses at least one of the combat algorithm and the z-score method. In this embodiment, the combat algorithm is first used to perform batch correction on the extracted eigenvalues, and then the data corrected by the combat algorithm is corrected using the z-score method. In this embodiment, the combat algorithm and the z-score method are both implemented by Python software.

[0059] S5. Given a threshold value of B% for efficacy. According to qualified data, for a patient, if the volume of the primary nasopharyngeal carcinoma tumor in the MRI image at the time of induction chemotherapy for n weeks is reduced by more than B% compared with the volume of the primary nasopharyngeal carcinoma tumor in the MRI image before induction chemotherapy, then the induction chemotherapy efficacy value of the patient is C. Otherwise, the induction chemotherapy efficacy value of the patient is D. Both C and D are given values, C>D.

[0060] Specifically, B%≥30%, C is 1, and D is 0. The larger n is, the larger B is. In this embodiment, B% is 30%.

[0061] Further, the analysis object data with the induction chemotherapy efficacy value of D is copied integer times, so that the ratio between the number of analysis object data with the induction chemotherapy efficacy value of D and the number of analysis object data with the induction chemotherapy efficacy value of C is closest to 1:1, so as to reduce the influence of the large difference in the number of analysis objects with different induction chemotherapy efficacy on the accuracy of subsequent operations. In this embodiment, the analysis objects with the induction chemotherapy efficacy value of D account for 21.7% of all analysis objects in S5, so in this embodiment, the analysis object data with the induction chemotherapy efficacy value of D is copied into three copies (one original data and two copies).

[0062] S6. Perform Mann-Whitney U test and Spearman correlation analysis according to the characteristic values ​​of the stable radiomics feature types, and screen out stable radiomics feature types with a P value less than 0.05 and a Spearman correlation coefficient greater than 0.6 in the Mann-Whitney U test results. In this embodiment, the Mann-Whitney U test and Spearman correlation analysis are both implemented by Python software.

[0063] S7. Use the glmnet function and cv.glmnet function in Python software to screen the stable radiomics feature types screened out in S6 based on the elastic network regression algorithm. The formula of the glmnet function is as follows:

[0064]

[0065] Where α is 1, λ is determined by the cv.glmnet function, w is the regression coefficient for the stable radiomics feature type, and x i w is the characteristic value of the i-th patient corresponding to the stable radiomics feature type, y i For x i The corresponding induction chemotherapy efficacy value of the patient, N is the total number of analysis subjects in the qualified data.

[0066] Calculate w for each stable radiomics feature type screened by S6. If a stable radiomics feature type does not have w, remove the stable radiomics feature type. If a stable radiomics feature type has w, retain the stable radiomics feature type as the radiomics feature type of nasopharyngeal carcinoma.

[0067] The types of nasopharyngeal carcinoma radiomics features selected in this embodiment involve kurtosis, cluster prominence, grayscale non-uniformity, cluster trend, dependency entropy, maximum correlation coefficient, contrast and intensity of T2WI, and minimum value, skewness, cluster shadow, grayscale non-uniformity, regional entropy and kurtosis of CET1WI.

[0068] There are 9 radiomic features of NPC on T2WI, namely kurtosis (square_firstorder_Kurtosis), cluster prominence (wavelet.LLH_glcm_ClusterProminence), gray level non-uniformity (wavelet.LLH_glszm_GrayLevelNon Uniformity), cluster tendency (wavelet.LHH_glcm_ClusterTendency), dependency entropy (wavelet.LHH_gldm_De pendenceEntropy), maximum correlation coefficient (wavelet.HLH_glcm_MCC), dependency entropy (wavelet.HLH_gldm_Dep endenceEntropy), contrast (wavelet.HLH_ngtdm_Contrast) and strength (wavelet.HHL_ngtdm_Strength).

[0069] The radiomic features of NPC in CET1WI include six features, namely, minimum (square_firstorder_Minimum), skewness (wavelet.LLH_firstorder_Skewness), cluster shadow (wavelet.LLH_glcm_ClusterShade), gray level non-uniformity (wavelet.LLH_glszm_GrayLevelNonUniformity), regional entropy (wavelet.HHH_glszm_ZoneEntr opy), and kurtosis (wavelet.LLL_firstorder_Kurtosis).

[0070] S8. Using the characteristic value of the radiomics feature type of nasopharyngeal carcinoma as the independent variable and the efficacy of induction chemotherapy for patients determined by S5 as the dependent variable, a GNN neural network model was trained to predict the efficacy of induction chemotherapy for patients with nasopharyngeal carcinoma.

[0071] The GNN neural network model of this embodiment uses ReLu as the activation function, the learning rate is set to 0.0001, and the number of hidden layer neurons is 24.

[0072] S9. Using the clinical pathological factors in S1 as independent variables and the efficacy of induction chemotherapy for patients determined by S5 as the dependent variable, univariate and multivariate regression analysis was performed to determine multiple clinical pathological factors with the greatest correlation with the efficacy of induction chemotherapy for patients as clinical pathological factors for nasopharyngeal carcinoma.

[0073] S10. Using the prediction results of the GNN neural network model in S8 and the data of clinical pathological factors of nasopharyngeal carcinoma in the clinical pathological data of S1 as independent variables, and the efficacy of induction chemotherapy for patients determined in S5 as the dependent variable, a machine learning model for predicting the efficacy of induction chemotherapy for nasopharyngeal carcinoma patients is trained, wherein the machine learning model is at least one of a GNN neural network model, an LR model, and a DT model.

[0074] The clinical pathological factors of nasopharyngeal carcinoma described in this example include hemoglobin concentration, neutrophil concentration and EBV-DNA concentration.

[0075] S10 of this embodiment trains a joint GNN model based on the GNN neural network model, S10 of this embodiment trains a joint LR model based on the LR model, and S10 of this embodiment trains a joint DT model based on the DT model.

[0076] Model interpretable analysis:

[0077] (1) In this example, the characteristic value of the radiomics feature type of nasopharyngeal carcinoma is used as the independent variable, and the efficacy of the patient's induction chemotherapy determined by S5 is used as the dependent variable. An LR model is constructed as a clinical LR model, and the clinical LR model is used to perform a comparative analysis with the GNN neural network model in S8.

[0078] This embodiment selects four samples for comparative analysis. The first sample is the training cohort, that is, the data sample used to train the clinical LR model and the GNN neural network model in S8; the second sample is the internal validation cohort, that is, the data sample used to evaluate the performance of the clinical LR model and the GNN neural network model in S8; the sample source of the training cohort and the internal validation cohort is the MRI image data from the 3.0T scanner of the same hospital; the third sample is the external validation cohort, which is used to evaluate the performance of the clinical LR model and the GNN neural network model in S8 in a real environment. The sample source of the external validation cohort is the MRI image data of the 3.0T scanner, and the samples of the training cohort and the internal validation cohort belong to different hospitals from the samples of the external validation cohort; the fourth sample is the cross-field strength (1.5T) validation cohort. This embodiment selects 291 analysis subjects from one of the three hospitals in S1. The MRI images in the data of these analysis subjects were obtained by a 1.5T scanner, and the data samples obtained based on the MRI images obtained by these 1.5T scanners are used as the cross-field strength (1.5T) validation cohort.

[0079] The steps of the control analysis are: based on the prediction results, the number of true positives (TP, the number of diseased samples correctly predicted by the model), the number of false positives (FP, the number of diseased samples incorrectly predicted by the model as diseased), the number of true negatives (TN, the number of non-diseased samples correctly predicted by the model), and the number of false positives (FN, the number of diseased samples incorrectly predicted by the model as non-diseased) are collected, and the accuracy, sensitivity, and specificity of the clinical LR model and the GNN neural network model in S8 are calculated respectively. Accuracy = (TP+TN) / (TP+FP+TN+FN), sensitivity = TP / (TP+FN), and specificity = TN / (TN+FP). The ROC curves of the clinical LR model and the GNN neural network model in S8 are plotted with specificity as the horizontal axis and sensitivity as the vertical axis, and the AUC values ​​of the two ROC curves are calculated respectively. The difference in AUC values ​​of the two ROC curves is compared by the DeLong test. The results of the control analysis are shown in the figure. Figure 3 As shown in the figure, the P value of the Delong test shows that one of the clinical LR model and the GNN neural network model in S8 has significantly better performance than the other model. Combined with AUC, accuracy, sensitivity and specificity, it can be determined that the model with better performance between the clinical LR model and the GNN neural network model in S8 is the GNN neural network model in S8.

[0080] (2) 100 NPC patients were randomly selected and evaluated for the efficacy of induction chemotherapy by junior physicians, senior physicians, and the GNN neural network model in S8. The 100 NPC patients were divided into two categories: patients with good efficacy and patients with poor efficacy. Junior physicians and senior physicians made the evaluation based on MRI imaging information and clinical pathology information. Junior physicians and senior physicians were unaware of the pathology results. After the evaluation, the junior physicians and senior physicians combined the GNN neural network model in S8 for re-evaluation. The efficacy evaluation criteria for junior physicians and senior physicians are as follows:

[0081] 1) Morphologically: whether there is tumor necrosis, including the primary tumor and metastatic lymph nodes;

[0082] 2) The number of metastatic lymph nodes is ≥4;

[0083] 3) Higher EBV-DNA concentration (EBV-DNA concentration ≥ 5500).

[0084] 1 point is assigned if 1 point is met, patients with more than 2 points are classified as those with poor efficacy, and patients with less than 2 points are classified as those with good efficacy. Based on the evaluation results of the above evaluation, the AUC value, accuracy, sensitivity, specificity and cutoff value of each evaluation are calculated respectively. The cutoff value of the GNN neural network model in S8 is the prediction probability limit value when it divides patients into two categories of good efficacy and poor efficacy. The cutoff value of junior doctors and senior doctors is the evaluation limit value when they divide patients into two categories of good efficacy and poor efficacy based on the above criteria.

[0085] Evaluation results such as Figure 4 As shown, it can be found that both junior and senior doctors have greatly improved their performance in judging patient efficacy with the assistance of the GNN neural network model in S8.

[0086] (3) The predicted probability of the patient by the GNN neural network model in S8 is used as the risk label of each patient, and the disease probability of the four samples in the model effect analysis (1) is predicted. The predicted probability cutoff value of 0.313 for the best prognosis is determined by the minimum P value method. Based on this cutoff value, the patients are divided into low-risk group and high-risk group. Figure 5 , Figure 6 , Figure 7 , Figure 8 As shown, the Kaplan-Meier curves of the four samples in the model effect analysis (1) show that the GNN neural network model in S8 can significantly distinguish the prognosis of nasopharyngeal carcinoma patients after induction chemotherapy, which proves the effectiveness and generalization of the GNN neural network model in S8.

[0087] (4) The joint GNN model, joint LR model, joint DT model and clinical LR model were used to perform performance comparison analysis using the four samples in the model effect analysis (1). The analysis items included AUC value, accuracy, sensitivity and specificity. The analysis results are shown in Figure 2. Fig. 9 As shown, it can be seen that the performance of the joint GNN model, the joint LR model, and the joint DT model are relatively close, and the performance of the joint GNN model, the joint LR model, and the joint DT model are all better than that of the clinical LR model.

[0088] (5) Based on the SHAP software, the joint GNN model, the joint LR model and the joint DT model were interpreted, and the contribution of the prediction results of the GNN neural network model in S8, hemoglobin concentration, neutrophil concentration and EBV-DNA concentration to the prediction results of the joint GNN model, the joint LR model and the joint DT model was quantified respectively. The analysis results of the SHAP software showed that the prediction results of the GNN neural network model in S8 had the greatest contribution to the prediction results among the joint GNN model, the joint LR model and the joint DT model.

[0089] The GNN neural network model in S8 was interpreted based on SHAP software, and the contribution of each NPC radiomics feature type to the prediction results of the GNN neural network model in S8 was quantified. Fig.10 In the middle are the top 10 contributing radiomic feature types of NPC. Fig.10 In the figure, red represents a higher eigenvalue, blue represents a lower eigenvalue, points on the left side of the Y axis increase the chance of a good therapeutic effect, and points on the right side of the Y axis reduce the chance of a good therapeutic effect. It can be seen that the top three NPC radiomics feature types of efficacy prediction, namely CET1WI kurtosis (wavelet.LLL_firstorder_Kurtosis), CET1WI gray level non-uniformity (wavelet.LLH_glszm_GrayLevelNonUniformity) and cluster trend (wavelet.LHH_glcm_ClusterTendency), are positively correlated with the size of the SHAP value, indicating that these three types are risk factors, which may contribute to the poor efficacy of induced chemotherapy in NPC patients.

[0090] In summary, this embodiment provides a clinical chemotherapy efficacy prediction and analysis method based on a multi-center MRI image interpretable model, by collecting MRI images of patients, screening and preprocessing the MRI images, extracting radiomic features, and based on the intraclass correlation coefficient ICC, Mann-Whitney U test, Spearman correlation analysis and elastic network regression algorithm, screening nasopharyngeal carcinoma radiomic features from radiomic features to construct a GNN neural network model to predict the induction chemotherapy response of patients with locally advanced nasopharyngeal carcinoma, which will help enhance the confidence of clinicians in treatment decisions, help clinicians determine the best time for treatment, and reduce the occurrence of over-medicalization. At the same time, this embodiment also establishes a machine learning model based on the prediction results of the GNN neural network model combined with clinical pathological factors, which further improves the prediction accuracy of the induction chemotherapy response of patients with locally advanced nasopharyngeal carcinoma.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting and analyzing the efficacy of clinical chemotherapy based on a multi-center MRI image interpretable model, characterized by: The steps include: S1. Select patients with nasopharyngeal carcinoma diagnosed by pathology in multiple hospitals as analysis subjects, and obtain data of the analysis subjects, the data including MRI images before induction chemotherapy and at n weeks after induction chemotherapy, where n is a given value; the MRI images include T2WI and CET1WI; S2, removing unqualified data, and processing the remaining MRI images using the SimpleITK package to obtain qualified data, wherein the processing includes N4 bias field correction and standardized voxel spacing; The removal criteria include:

1. For nasopharyngeal carcinoma primary tumors that cannot be measured in MRI images before induction chemotherapy but are determined based on pathological analysis, the MRI images of the location of the primary nasopharyngeal carcinoma tumor are removed; 2. MRI images that cannot be analyzed by imaging genomics are removed; 3. Analysis object data without MRI images before induction chemotherapy are removed; 4. Analysis object data that have not received complete induction chemotherapy are removed; 5. Analysis object data with a history of anti-tumor treatment before the first MRI examination are removed; 6. Analysis object data with existing or concurrent other primary malignant tumors are removed; S3. Randomly select at least 50 MRI images before induction chemotherapy from the qualified data as stability assessment data; select multiple radiologists, and each radiologist independently segment the stability assessment data to segment the region of interest of nasopharyngeal carcinoma tumors; use the Pyradiomics package to extract radiomic feature data from the region of interest segmented by each radiologist. The radiomic feature data includes the type and the characteristic value corresponding to each type, and the intraclass correlation coefficient ICC is calculated for each extracted radiomic feature type; given a stability threshold A, the radiomic feature type with an ICC greater than A is regarded as a stable radiomic feature type; S4. Multiple radiologists discussed and processed the MRI images before induction chemotherapy in the qualified materials, segmented the regions of interest of nasopharyngeal carcinoma tumors, and used the Pyradiomics package to extract feature values ​​of stable radiomics feature types from the regions of interest segmented by multiple radiologists; S5. Given a threshold value of efficacy B%; based on qualified data, for a patient, if the volume of the primary nasopharyngeal carcinoma tumor in the MRI image at the time of induction chemotherapy for n weeks is reduced by more than B% compared with the volume of the primary nasopharyngeal carcinoma tumor in the MRI image before induction chemotherapy, then the induction chemotherapy efficacy value of the patient is C, otherwise the induction chemotherapy efficacy value of the patient is D; C and D are both given values, C>D; S6. Perform Mann-Whitney U test and Spearman correlation analysis based on the characteristic values ​​of stable radiomics feature types, and screen out stable radiomics feature types with a P value less than 0.05 and a Spearman correlation coefficient greater than 0.6 in the Mann-Whitney U test results; S7. Use the glmnet function and cv.glmnet function to screen the stable radiomics feature types screened out in S6 based on the elastic network regression algorithm. The formula of the glmnet function is as follows: Where α is 1, λ is determined by the cv.glmnet function, w is the regression coefficient for the stable radiomics feature type, and x i w is the characteristic value of the i-th patient corresponding to the stable radiomics feature type, y i For x i The corresponding patient's induction chemotherapy efficacy value, N is the total number of analysis subjects in the qualified data; Calculate w for each stable radiomics feature type screened by S6. If a stable radiomics feature type does not have w, remove the stable radiomics feature type. If a stable radiomics feature type does have w, retain the stable radiomics feature type as the radiomics feature type of nasopharyngeal carcinoma. S8. Using the characteristic value of the radiomics feature type of nasopharyngeal carcinoma as the independent variable and the efficacy of induction chemotherapy for patients determined by S5 as the dependent variable, a GNN neural network model was trained to predict the efficacy of induction chemotherapy for patients with nasopharyngeal carcinoma.

2. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 1, characterized in that: In S3, A≥0.

75.

3. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 2, characterized in that: In S1, n≥2, and in S5, B%≥30%.

4. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 3, characterized in that: In S5, C is 1 and D is 0.

5. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 4, characterized in that: The types of radiomic features of NPC included kurtosis, cluster prominence, grayscale heterogeneity, cluster trend, dependency entropy, maximum correlation coefficient, contrast and intensity of T2WI, and minimum value, skewness, cluster shadow, grayscale heterogeneity, regional entropy and kurtosis of CET1WI.

6. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 1, characterized in that: S5 also includes duplicating the analysis object data with the induction chemotherapy efficacy value D by an integer multiple so that the ratio between the number of analysis object data with the induction chemotherapy efficacy value D and the number of analysis object data with the induction chemotherapy efficacy value C is closest to 1:

1.

7. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 1, characterized in that: S4 also includes batch correction of the extracted feature values ​​to remove the batch effects caused by different hospitals and different scanners.

8. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 7, characterized in that: In S4, batch correction uses at least one of the combat algorithm and the z-score method.

9. A method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to any one of claims 1 to 8, characterized in that: The data described in S1 also include clinical pathological data before induction chemotherapy and at n weeks after induction chemotherapy. The clinical pathological factors in the clinical pathological data include age, gender, smoking history, family history, weight, height, body mass index, hemoglobin concentration, white blood cell concentration, neutrophil concentration, lymphocyte concentration, neutrophil-lymphocyte ratio, platelet concentration, albumin concentration, EBV-DNA concentration, clinical T stage, clinical N stage and WHO type; The removal criteria for S2 also include 7. Removal of data of subjects without clinical pathological data before induction chemotherapy; The following steps are also included: S9, using the clinical pathological factors in S1 as independent variables and the efficacy of induction chemotherapy determined by S5 as the dependent variable, univariate and multivariate regression analysis was performed to determine multiple clinical pathological factors with the greatest correlation with the efficacy of induction chemotherapy as clinical pathological factors of nasopharyngeal carcinoma; S10. Using the prediction results of the GNN neural network model in S8 and the data of clinical pathological factors of nasopharyngeal carcinoma in the clinical pathological data of S1 as independent variables, and the efficacy of induction chemotherapy for patients determined in S5 as the dependent variable, a machine learning model for predicting the efficacy of induction chemotherapy for nasopharyngeal carcinoma patients is trained, wherein the machine learning model is at least one of a GNN neural network model, an LR model, and a DT model.

10. The method for predicting and analyzing clinical chemotherapy efficacy based on a multi-center MRI image interpretable model according to claim 9, characterized in that: The clinical pathological factors of nasopharyngeal carcinoma include hemoglobin concentration, neutrophil concentration and EBV-DNA concentration.

Images