A method, system, device and storage medium for processing radiomics data of recurrent nasopharyngeal carcinoma

Abstract

The present invention discloses a method, system, device and storage medium for processing radiomics data of recurrent nasopharyngeal carcinoma. The radiomics data processing method comprises: obtaining MRI image data of patients with recurrent nasopharyngeal carcinoma; extracting image data for analysis, segmenting lesions, extracting image features, screening radiomics key features; inputting radiomics key features into the random forest model obtained through training to predict the risk of nasopharyngeal necrosis. The present invention provides accurate guidance for the prediction of the risk of nasopharyngeal necrosis and the formulation of treatment plans for patients with recurrent nasopharyngeal carcinoma, so that a part of patients with high risk of nasopharyngeal necrosis can avoid adverse outcomes after re-radiotherapy, and promote individualized and accurate diagnosis and treatment of recurrent nasopharyngeal carcinoma.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a method, system, device and storage medium for processing recurrent nasopharyngeal carcinoma imaging genomics data. Background Art

[0002] Nasopharyngeal carcinoma is a malignant tumor of the head and neck, and radiotherapy is the main treatment for nasopharyngeal carcinoma. In recent years, due to the development of intensity-modulated conformal radiotherapy technology, the local control rate of nasopharyngeal carcinoma has been significantly improved, but 10% to 15% of patients still relapse. Re-radiotherapy is currently the most important and effective treatment for recurrent nasopharyngeal carcinoma. Studies have shown that patients with recurrent nasopharyngeal carcinoma may still achieve better local control (3-year local control rate is 66.1%-71.0%) and long-term survival (3-year overall survival is 44.3%-51.8%) after receiving re-radiotherapy.

[0003] However, the serious complications caused by re-radiotherapy are an important problem that troubles clinicians. Among them, nasopharyngeal necrosis is the most common complication after re-radiotherapy, which has been fully reported in previous literature. Nasopharyngeal necrosis destroys the defense environment of the normal nasopharyngeal mucosa and surrounding soft tissues, resulting in the exposure of important large blood vessels and skull base bones, and is very easy to breed bacterial infection, causing serious complications such as intracranial infection and nasopharyngeal hemorrhage, which is an important cause of death in patients undergoing re-radiotherapy. Previous literature reports that the incidence of nasopharyngeal necrosis in patients receiving re-radiotherapy is about 28.8% to 50.8%, which is significantly higher than that in patients receiving primary radiotherapy (incidence of about 2.1%). If the nasopharyngeal tolerance of re-radiotherapy can be fully evaluated before treatment, a reliable nasopharyngeal necrosis prediction model can be established, and a part of high-risk patients with nasopharyngeal necrosis who cannot tolerate re-radiotherapy can be screened out, it will help guide the clinical individualization of re-radiotherapy for recurrent nasopharyngeal carcinoma, and it can also intervene early in the high-risk population for necrosis, thereby avoiding adverse outcomes after nasopharyngeal necrosis. Therefore, developing a scientific and effective nasopharyngeal necrosis prediction model is of great significance for further improving the survival prognosis of patients with recurrent nasopharyngeal carcinoma.

[0004] In the past, a few studies have constructed nasopharyngeal necrosis prediction models based on simple clinical factors, but the results were not ideal. In 2016, Yu et al. found in 228 patients with recurrent nasopharyngeal carcinoma who received re-radiotherapy that female sex, necrosis before re-radiotherapy, higher cumulative dose of two radiotherapy sessions (≥141.5Gy), and larger recurrent tumor volume (≥25.38cm3) were high-risk factors for fatal nasopharyngeal necrosis, but failed to establish a reliable individualized prediction model. In 2019, Li et al. included a large sample cohort of 7144 nasopharyngeal carcinoma patients and constructed a multivariate regression model based on age, pathological type, history of diabetes, T stage, radiotherapy technology, whether to re-radiotherapy, and blood indicators (C-reactive protein, hemoglobin, albumin levels). Although the prediction accuracy of the model reached 0.78, the vast majority of the included patients had received primary radiotherapy multiple times (only 3.4% of patients received re-radiotherapy), and clinical application was insufficient. We collected 625 patients with recurrent nasopharyngeal carcinoma who received re-radiotherapy at the Sun Yat-sen University Cancer Center and established a comprehensive clinical database. The results showed that the best prediction accuracy of the clinical model for nasopharyngeal necrosis after re-radiotherapy was only 0.61. The above results show that clinical information has limited effectiveness in predicting nasopharyngeal necrosis, and it is urgent to use more abundant information and advanced data mining methods to further improve the prediction effect of the model.

[0005] Imaging examination (especially MRI scan of the nasopharynx) is the most important means in the diagnosis and treatment of NPC. In fact, in addition to clearly showing the morphological characteristics of the lesions, MRI images also contain a series of digital information that can be deeply mined. In 2014, Mitra S defined "radiomic" as: high-throughput and automatic extraction of a large amount of radiomics data from the area of ​​interest of medical images, converting it into high-resolution, mineable spatial data, and then performing quantitative analysis to obtain target information, comprehensively evaluate various phenotypes such as lesion pathology and genetics, and assist in making clinical decisions. As a product of the intersection of medicine and engineering, radiomics can quantitatively decode target tissue, cell and subcellular level information in vivo, and is a research hotspot in the medical field. In recent years, radiomics technology based on MRI or PET / CT has been used to predict the prognosis of NPC and the efficacy of induction chemotherapy, and has achieved ideal results. Radiomics technology provides new opportunities for the evaluation of nasopharyngeal tolerance to re-radiotherapy. Summary of the invention

[0006] The purpose of the first aspect of the present invention is to provide a marker for predicting the risk of nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy.

[0007] The purpose of the second aspect of the present invention is to provide an application of the substance for detecting the marker described in the first aspect of the present invention.

[0008] The third aspect of the present invention aims to provide a product.

[0009] The fourth aspect of the present invention aims to provide a method for constructing a risk prediction model for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy.

[0010] The fifth aspect of the present invention aims to provide a risk prediction system for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy.

[0011] A sixth aspect of the present invention aims to provide an electronic device.

[0012] The seventh aspect of the present invention aims to provide a storage medium.

[0013] The technical solution adopted by the present invention is:

[0014] According to a first aspect of the present invention, a marker for predicting the risk of nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy is provided, comprising: log-sigma-3-0-mm-3D_firstorder_RootMeanSquared_T2, original_shape_Maximu m2DDiameterColumn_T1, original_shape_Sphericity_T1C, wavelet-HL_glcm_Imc2_T1C, log-sigma-1-0-mm-3D_glcm_ClusterProminence_T2, and wavelet-HL_glrlm_RunEntropy_T1.

[0015] The second aspect of the present invention provides the use of a substance for detecting the marker described in the first aspect of the present invention in the preparation of a product for predicting the risk of nasopharyngeal necrosis after re-radiotherapy in patients with recurrent nasopharyngeal carcinoma.

[0016] The third aspect of the present invention provides a product, wherein the product comprises a substance for detecting a marker, and the marker is as described in the first aspect of the present invention.

[0017] A fourth aspect of the present invention provides a method for constructing a risk prediction model for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy, comprising the following steps:

[0018] Obtain MRI imaging data of patients with recurrent nasopharyngeal carcinoma;

[0019] Extract image data for analysis, segment lesions, extract image features, and screen key features of imaging omics;

[0020] Machine learning algorithms were used to establish a model based on key features of radiomics to predict the risk of nasopharyngeal necrosis.

[0021] In some embodiments of the present invention, the radiomics key feature is the marker described in the first aspect of the present invention.

[0022] In some embodiments of the present invention, the specific steps of screening the key features affecting omics are:

[0023] (1) Repeatability test: Based on repeated segmentation data, the inter-class / intra-class correlation coefficient (ICC) was used to evaluate the repeatability of radiomics features, and stable features with ICC>0.7 were selected;

[0024] (2) Importance test: The support vector machine (SVM) binary classifier was used to rank the ability of each feature to distinguish between cases with nasopharyngeal necrosis and cases without nasopharyngeal necrosis, and the 100 most important features in each sequence were retained;

[0025] (3) Feature elimination and independence test: The random forest algorithm was used to construct the radiomics model. The recursive feature elimination method was used to filter the features according to the accuracy of the model in predicting nasopharyngeal necrosis. The grid search strategy was used to determine the optimal number of features to be retained. The inter-feature correlation coefficient threshold was set to 0.7 in the recursive feature elimination method, and the radiomics features with a correlation coefficient < 0.7 were retained.

[0026] In some embodiments of the present invention, the machine learning algorithm is a random forest, because the gradient boosting tree is prone to overfitting on a small data set; the extreme random tree does not perform sample randomization and is also prone to overfitting on a small data set.

[0027] According to a fifth aspect of the present invention, there is provided a risk prediction system for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy: the system is established by the construction method described in the fourth aspect of the present invention.

[0028] In a sixth aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for constructing a risk prediction model for nasopharyngeal necrosis occurring in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy as described in the fifth aspect of the present invention is implemented.

[0029] According to a seventh aspect of the present invention, a storage medium is provided, in which processor-executable instructions are stored. When the processor-executable instructions are executed by the processor, they are used to execute the method described in the fourth aspect of the present invention.

[0030] The present invention also provides an application of imaging features in constructing a risk prediction model for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy, wherein the imaging features are as described in the first aspect of the present invention.

[0031] The beneficial effects of the present invention are:

[0032] The imaging genomics model for predicting nasopharyngeal necrosis after re-radiotherapy based on multi-sequence MRI images provided by the present invention is a non-invasive means that can accurately evaluate the nasopharyngeal tolerance of patients with recurrent nasopharyngeal carcinoma to re-radiotherapy before treatment, and can accurately predict the risk of nasopharyngeal necrosis after re-radiotherapy for each patient based on MRI images before treatment. For patients at high risk of nasopharyngeal necrosis, clinicians should carefully choose re-radiotherapy, and can give priority to palliative chemotherapy, immunotherapy and targeted therapy; while patients with low risk of nasopharyngeal necrosis may be the best population for re-radiotherapy. Therefore, the imaging genomics model can guide the clinical individualization of re-radiotherapy for recurrent nasopharyngeal carcinoma, which is of great significance for guiding clinical practice. The present invention is expected to make a breakthrough in the assessment of tolerance to re-radiotherapy, provide accurate guidance for the prediction of the risk of nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma and the formulation of treatment plans, so that a part of patients with high risk of nasopharyngeal necrosis can avoid adverse outcomes after re-radiotherapy, and promote individualized and accurate diagnosis and treatment of recurrent nasopharyngeal carcinoma.

[0033] The method provided by the present invention constructs an accurate nasopharyngeal necrosis prediction model based on high-throughput imaging genomics features, applies machine learning algorithms such as support vector machines, recursive feature elimination, and random forests, and effectively solves the problems of high throughput, multiple redundancy, intra-group variation, and inter-group ambiguity of imaging genomics features. It is also verified and optimized on a multi-center cohort to improve the prediction accuracy of the model, and the constructed model has good prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The figure is a flow chart of the method of the present invention.

[0035] Figure 2 Contribution graph for model features.

[0036] Figure 3 It is the ROC curve diagram.

[0037] Figure 4 This is the model calibration curve. DETAILED DESCRIPTION

[0038] The following will be combined with the embodiments to clearly and completely describe the concept of the present invention and the technical effects produced, so as to fully understand the purpose, characteristics and effects of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0039] Example 1

[0040] Technical flow chart see Figure 1 , the specific implementation is as follows:

[0041] 1. Establish a multicenter study cohort of patients with recurrent NPC undergoing re-radiation

[0042] A total of 625 patients were collected from the Cancer Center of Sun Yat-sen University and randomly divided into a training set (420 cases) and an internal validation set (205 cases) in a ratio of 7:3; patients were collected from three medical centers, namely, the Cancer Hospital of the Chinese Academy of Medical Sciences, the Affiliated Cancer Hospital of Fudan University, and the Nanfang Hospital of Southern Medical University, to form an external validation set (136 cases). All patients were pathologically diagnosed with locally recurrent NPC and received a radical dose (≥60 Gy) of re-radiotherapy (intensity-modulated or helical tomotherapy). Pre-treatment MRI images of patients, as well as clinical and follow-up data, including age, gender, body mass index (BMI), interval between two radiotherapy sessions, baseline hemoglobin level, re-radiotherapy dose, recurrent tumor volume, nasopharyngeal necrosis, and overall survival (OS) were collected.

[0043] 2. MRI image acquisition and lesion segmentation

[0044] The original MRI images (including T1, T2, and T1C sequences) were exported from a medical scanning machine, and experienced radiotherapists used ITK-SNAP software to outline the nasopharyngeal tumor lesions layer by layer in the cross-sections of the three sequences. Repeated segmentation data was established for 50 of the cases: another different researcher and the same researcher repeated the segmentation of the cases four weeks later to test the inter-observer and intra-observer stability of the features.

[0045] 3. Radiomics feature extraction

[0046] All images were resampled to a voxel resolution of 0.5 mm × 0.5 mm × 3.0 mm using a linear interpolation method. The pyradiomics package of Python software was used to extract radiomic features on the original image data and the enhanced data after various filters. These features included five types: first-order statistics, shape, texture (gray-level co-occurrence matrix [GLCM], gray-level dependency matrix [GLDM], gray-level run length matrix [GLRLM], and gray-level size zone matrix [GLSZM]), wavelet, and Laplace of Gaussian (LoG) transform features.

[0047] There are a total of 2106 radiomics features, including 602 for each of the T1, T2, and T1C sequences.

[0048] 4. Screening of radiomics features

[0049] In the training set, feature screening is performed in several steps:

[0050] (1) Repeatability test: Based on repeated segmentation data, the repeatability of radiomics features was evaluated using the inter- and intra-class correlation coefficients (ICC), and only stable features with ICC>0.7 were selected. There were a total of 1,630 radiomics features, including 503 T1 sequences, 550 T2 sequences, and 577 T1C sequences.

[0051] (2) Importance test: The support vector machine (SVM) binary classifier was used to rank the ability of each feature to distinguish cases with nasopharyngeal necrosis from cases without nasopharyngeal necrosis, and only the 100 most important features in each sequence were retained. There were a total of 300 radiomics features, including 100 for each of the T1, T2, and T1C sequences.

[0052] (3) Feature elimination and independence test: A radiomics model was constructed using a random forest algorithm. Recursive feature elimination (RFE) was used to filter features based on the accuracy of the model in predicting nasopharyngeal necrosis. A grid search strategy was used to determine the optimal number of features to be retained. At the same time, in order to address the multidimensionality and possible redundancy of radiomics features, the inter-feature correlation coefficient threshold was set to 0.7 in RFE, and only radiomics features with strong independence and a correlation coefficient < 0.7 were retained. There were a total of 6 radiomics features, including 2 for each of the T1, T2, and T1C sequences.

[0053] 5. Establishment of radiomics labels

[0054] A random forest model was constructed based on the six key radiomics features screened out, including log-sigma-3-0-mm-3D_firstorder_RootMeanSquared_T2, original_shape_Maximum2DDiameterColumn_T1, original_shape_Sphericity_T1C, wavelet-HL_glcm_Imc2_T1C, log-sigma-1-0-mm-3D_glcm_Clus terProminence_T2, and wavelet-HL_glrlm_RunEntropy_T1.

[0055] In the training set, the grid search strategy was first used to select the best model hyperparameters, where the search results for the number of tree models in the range of 0 to ∞ were 500, the search results for the sample downsampling rate in the range of 0 to 1 were 75%, the search results for the feature downsampling rate in the range of 0 to 1 were 80%, and the search results for the tree depth in the range of 1 to ∞ were 2. In the training set, the random forest model will simultaneously build 500 independent decision tree models, and take the average of their respective outputs as the output of the random forest model, which can increase the generalization of the random forest model. For each decision tree model, 75% of the total training samples were first randomly selected as the target samples of the tree model, and 80% of the features were randomly selected as the features for building the tree model. In order to reduce the overfitting of the tree model, the tree depth of a single decision tree model was limited to 2. The probability of nasopharyngeal necrosis predicted by the radiomics model was output for each patient in the training set and the validation set.

[0056] The contributions of the six radiomics features in the constructed random forest model are as follows: Figure 2 .

[0057] 6. Evaluation of model prediction accuracy

[0058] The model was constructed, internally validated, and externally tested for the training set, internal validation set, and external validation set. The receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) were used to evaluate the discrimination of the model in predicting nasopharyngeal necrosis. The calibration curve was drawn to evaluate the calibration of the radiomics model, and the Hosmer-Lemeshow goodness-of-fit test was used to quantitatively evaluate the model calibration.

[0059] The ROC curves of the radiomics model for predicting nasopharyngeal necrosis on the training set, internal validation set, and external validation set are shown in Figure 2. Figure 3 The AUC values ​​were 0.722 (95% CI 0.676-0.765), 0.713 (95% CI 0.653-0.772) and 0.756 (95% CI: 0.673-0.838), all above 0.7, indicating that the model prediction accuracy is high. The calibration of the radiomics model for predicting nasopharyngeal necrosis on the training set, internal validation set and external validation set is shown in Figure 2. Figure 4 , the P values ​​of the Hosmer-Lemeshow goodness-of-fit test were 0.584, 0.574 and 0.137, respectively, all greater than 0.05, indicating that the model passed the goodness-of-fit test, there was no significant difference between the predicted value and the true value of the model, and the consistency was good.

[0060] 7. Analysis of the prognostic value of imaging omics models

[0061] According to the median value of the radiomics score in the training set (0.735), patients were divided into high-risk and low-risk groups for nasopharyngeal necrosis. A multivariate Cox regression model adjusted for sex, age, BMI, interval between two radiotherapy treatments, baseline hemoglobin level, re-radiotherapy dose, and recurrent tumor volume was used to analyze whether the radiomics signature was an independent prognostic factor for OS.

[0062] The OS of patients with high risk of nasopharyngeal necrosis identified by the imaging omics model in the training set, internal validation set and external validation set was significantly worse than that of patients with low risk of nasopharyngeal necrosis, with death hazard ratios (HR) of 1.44 (95% CI 1.05-1.98), 2.04 (1.31-3.17) and 2.59 (1.33-5.07), respectively, indicating that the imaging omics model has important prognostic value.

[0063] The above specific implementations have been described in detail for the present invention, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the purpose of the present invention. In addition, the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

Claims

1. A method for constructing a risk prediction model for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy. It is characterized in that The following steps are involved: Obtain MRI imaging data of patients with recurrent nasopharyngeal carcinoma; Extract image data for analysis, segment lesions, extract image features, and screen key features of imaging omics; Machine learning algorithms were used to build models based on key radiomics features to predict the risk of nasopharyngeal necrosis; The key imaging features are: log-sigma-3-0-mm-3D_firstorder_RootMeanSquared_T2, original_shape_Maximum2DDiameterColumn_T1, original_shape_Spherici ty_T1C, wavelet-HL_glcm_Imc2_T1C, log-sigma-1-0-mm-3D_glcm_ClusterPromi nence_T2 and wavelet-HL_glrlm_RunEntropy_T1.

2. The method according to claim 1, It is characterized in that The specific steps for screening key features affecting omics are: (1) Repeatability test: Based on repeated segmentation data, the repeatability of radiomics features was evaluated using inter-group / intra-group correlation coefficients, and stable features with ICC>0.7 were selected; (2) Importance test: Use support vector machine binary classifier to rank the ability of each feature to distinguish cases with nasopharyngeal necrosis from cases without nasopharyngeal necrosis, and retain the 100 most important features in each sequence; (3) Feature elimination and independence test: The random forest algorithm was used to construct the radiomics model. The recursive feature elimination method was used to filter the features according to the accuracy of the model in predicting nasopharyngeal necrosis. The grid search strategy was used to determine the optimal number of features to be retained. The inter-feature correlation coefficient threshold was set to 0.7 in the recursive feature elimination method, and the radiomics features with a correlation coefficient < 0.7 were retained.

3. The method according to claim 1, It is characterized in that The machine learning algorithm is random forest.

4. A risk prediction system for nasopharyngeal necrosis in patients with recurrent nasopharyngeal carcinoma after re-radiotherapy. It is characterized in that The method is established by the construction method according to any one of claims 1 to 3.

Images