Prediction model for predicting liver cancer combined treatment effect based on multi-mode MRI radiomics characteristics and construction method and application thereof

Abstract

The invention discloses a prediction model for predicting a liver cancer combined treatment effect based on multi-mode MRI radiomics characteristics and a construction method and application thereof, and belongs to the technical field of liver cancer treatment effect prediction. In order to solve the problem that only part of liver cancer patients can benefit from an arterial chemoembolization combined targeted immunotherapy scheme at present, the invention proposes that before combined therapy, MRI radiomics characteristics and clinical factors are enhanced through noninvasive multi-modal comparison, multiple machine learning models are established, and the model learning model is used for learning the liver cancer patients. The tumor treatment response of arterial chemoembolism combined with molecular targeting and immunotherapy is predicted, and the applicability of the radiomics model is verified by using an external data set; besides, the radiomics model is combined with a clinical model, so that the method has more excellent prediction efficiency, and the radiomics features provide incremental prediction value for clinical factors. The prediction model can predict the combined treatment effect of the liver cancer patients, so that corresponding more effective and reasonable treatment schemes are selected for different patients before treatment.

Description

Technical Field

[0001] The present invention relates to the technical field of liver cancer treatment effect prediction, and in particular to a prediction model for predicting liver cancer combined treatment effects based on multimodal MRI radiomics features, and a construction method and application thereof. Background Art

[0002] Hepatocellular carcinoma (HCC) is one of the most common malignant tumors and the second leading cause of cancer-related death. Most patients are already in the middle or late stages of the tumor when diagnosed and have lost the opportunity for surgical resection. Transarterial chemoembolization (TACE) is the most commonly used local treatment for HCC that cannot be surgically removed. In recent years, immunotherapy represented by PD-(L)1 (Programmed cell death1and ligand) immune checkpoint inhibitors has pushed the systemic treatment of HCC into a new era. Targeted immunotherapy based on molecular targeted therapy (MTT) and immunotherapy has become the first-line systemic treatment for advanced HCC.

[0003] Combination therapy of different treatment methods, local therapy and systemic therapy, is the key to improving the transformation rate of advanced HCC and improving patient survival. The purpose is to explore treatments with higher objective response rates (ORR). Local therapy and targeted immunotherapy have synergistic effects, which can further improve the therapeutic efficacy. Some important retrospective studies have shown that compared with single TACE treatment, TACE combined with targeted immunotherapy has better efficacy and a higher ORR. However, in clinical practice, only some patients benefit from combined therapy. Therefore, in the era of precision medicine, there is an urgent need to explore effective predictive factors and markers before treatment, making them a useful tool for predicting the response of HCC patients to combined therapy, so as to classify HCC patients before treatment, make reasonable clinical decisions, and guide patients to receive appropriate treatment.

[0004] Radiomics is a rapidly developing research field in recent years. It aims to extract a large amount of information from imaging images with high throughput, convert images of regions of interest into quantitative data, and realize the extraction of lesion features and the establishment of models. Magnetic resonance imaging (MRI) provides higher soft tissue contrast resolution than computed tomography (CT) images, and contrast-enhanced MRI remains the most reliable radiological technique for evaluating the response to systemic treatment of liver cancer. Radiomic features extracted from MRI images have provided important information and support for the histological grade, microvascular invasion, surgery, local TACE treatment, systemic treatment response, and prognosis of HCC. However, there are currently no available multimodal MRI radiomics studies evaluating the response of HCC patients to TACE combined with MTT plus immunotherapy. Summary of the invention

[0005] In response to the above-mentioned problems, the present invention aims to provide a prediction model based on multimodal MRI radiomics features for predicting the effect of combined treatment of liver cancer, as well as a construction method and application thereof. Through non-invasive multimodal contrast-enhanced MRI radiomics features before treatment, multiple machine learning models are established to predict the tumor treatment response of TACE combined with MTT plus immunotherapy, and the applicability of the radiomics model is verified using a multicenter external data set.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features, wherein the nomogram of the prediction model includes:

[0008] The first line is a score scale, ranging from 0 to 100;

[0009] The second row shows the alpha-fetoprotein AFP values, including ≤400 and >45;

[0010] The third row is the neutrophil-to-lymphocyte ratio (NLR), ranging from 0 to 11;

[0011] The fourth line is the radiomics score Rad-score, ranging from -1 to 0.6;

[0012] The fifth line is the total score, which is the sum of the numbers in the second to fourth lines, ranging from 0 to 140;

[0013] The sixth line is the linear prediction value, ranging from -12 to 8;

[0014] The seventh line is the prediction value of treatment effect, ranging from 0.1 to 0.95.

[0015] Furthermore, the present invention also includes a method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features as described above, comprising the following steps:

[0016] S1: Obtain contrast-enhanced MRI images of the abdomen of patients with liver cancer before starting combined therapy;

[0017] S2: Segment the contrast-enhanced abdominal MRI images of patients to obtain radiomic features;

[0018] S3: Selection of clinically independent predictors and radiomic features for building a prediction model;

[0019] S4: Use different machine learning methods to build clinical models and radiomics models respectively;

[0020] S5: Perform model evaluation on the model constructed in step S4 to determine the best clinical model and radiomics model;

[0021] S6: combining the best clinical model and radiomics model determined in step S5 to obtain a joint model, i.e., the final prediction model;

[0022] S7: Conduct external validation of the constructed joint model.

[0023] Furthermore, the patient's abdominal MRI contrast-enhanced image described in step S1 includes six modality images, namely, T1WI, T2WI, DWI, arterial phase, portal venous phase, and delayed phase.

[0024] Furthermore, the clinical independent predictive factors described in step S3 include the neutrophil-to-lymphocyte ratio NLR and the alpha-fetoprotein AFP value.

[0025] Further, the radiomics features described in step S3 include Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_AP, Original_glrlm ShortRunHighGrayLevelEmphasis_AP, Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_PP, Wavelet-LLL_ngtdm_Strength_AP, Gradient_firstorder_10Percentile_AP, Wavelet-HLL_firstorder_RootMeanSquared_AP, Wavelet-LL_gldm_DependenceVariance_DP, Exponential_gldm_DependenceVariance_DP, Original_g ldm_DependenceVariance_DWI, Wavelet-LHL_gldm_DependenceEntropy_T2, lbp-3D-k_glrlm_ShortRunEmphasis_DP, Wavelet-LL_glrlm_ShortRunHighGrayLevelE mphasis_DP, square_firstorder_Energy_DWI, Wavelet-LHL_glszm_SmallAreaEmphasis_PP, Wavelet-LHH_ngtdm_Complexity_T1 and Original_shape_Elongation_T1.

[0026] Furthermore, the machine learning method described in step S4 includes logistic regression, K-nearest neighbor, support vector machine, gradient boosting, decision tree and random forest.

[0027] Furthermore, the optimal clinical model and radiomics model described in step S5 are clinical model and radiomics model of logistic regression.

[0028] Furthermore, the present invention also includes the application of the prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features as described above in the selection of treatment plans before treatment of liver cancer patients.

[0029] The beneficial effects of the present invention are:

[0030] 1. In the present invention, when constructing a prediction model based on multimodal MRI radiomics features to predict the effect of combined treatment of liver cancer, 6 different modalities of MRI images were used, namely T1 weighted sequence (T1WI), T2 weighted sequence (T2WI), diffusion weighted sequence (DWI), arterial phase (AP), portal venous phase (PP), delayed phase (DP) sequence. Compared with previous MRI radiomics studies that usually only used 1-2 enhanced single modalities, more modalities can provide more information characteristics of different tumor modalities, which can reflect tumor treatment response from different dimensions. These multimodal high-dimensional features contain more sophisticated tumor biological information and more sensitive treatment response predictions.

[0031] 2. The present invention adopts a variety of different machine learning models for modeling. Since the effectiveness of various prediction models is different, the accuracy of the prediction model can be improved by selecting the best prediction model from a variety of different machine learning models.

[0032] 3. In the present invention, 16 radiomic features were finally screened from multiple radiomic features of 6 different modalities for constructing a prediction model. The clinical decision-making value and incremental predictive value results showed that radiomics can effectively identify the potential and predictive value of unresectable HCC patients who are most likely to respond to TACE combined with targeted immunotherapy. The radiomic features of multimodal MRI provide incremental predictive value relative to traditional clinical factors and can be widely used to help clinicians make treatment decisions based on TACE combined with targeted immunotherapy for HCC patients, and the combined model of radiomics model and clinical model has better prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a diagram showing the results of patient screening and grouping in the present invention;

[0034] Figure 2 This is an example diagram of segmentation based on six modes of tumor MRI images in the present invention;

[0035] Figure 3 A flowchart for the development and validation of the prediction model in the present invention;

[0036] Figure 4 Receiver operating characteristic curves of the clinical model of the present invention in three cohorts;

[0037] Figure 5 is the selection result of the radiology group features in the present invention;

[0038] Figure 6 It is the heat map expression of 16 radiomic features in the present invention in three arrays;

[0039] Figure 7Rad score in the present invention and its relationship with tumor treatment response;

[0040] Figure 8 is the receiver operating characteristic curve of the radiomics model and the combined model based on machine learning in the present invention;

[0041] Fig. 9 It is the nomogram expression of the joint model in the present invention;

[0042] Fig.10 Radar charts of the clinical model, radiomics model and joint model and confusion matrix of the joint model in the present invention;

[0043] Fig.11 These are the calibration curve and decision curve of the joint model in the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0045] 1 Research subjects

[0046] A total of 276 patients with advanced HCC from four centers from January 2019 to June 2023 were included in this study, and all patients were divided into three research cohorts: the Cancer Hospital of the Chinese Academy of Medical Sciences (center 1) was used as the training cohort (N=152), and its data were used to establish the model; the First Affiliated Hospital of Soochow University (center 2) was used as the external validation cohort 1 (N=60); Guizhou Cancer Hospital (center 3) (N=59) and Jilin University China-Japan Friendship Hospital (center 4) (N=5) were used as the external validation cohort 2 (N=64), and the patient screening and grouping were as shown in the attached Figure 1 shown.

[0047] All patients underwent contrast-enhanced MRI within 1 month before the start of combined therapy, and tumor response was assessed every 60 days (±7 days) according to mRECIST criteria. Inclusion criteria included: (1) age ≥ 18 years; (2) histologically or clinically confirmed Barcelona Clinical Cancer (BCLC) stage B or C HCC; (3) Child-Pugh grade A or B; (4) ECOG performance status 0 or 1; (5) concurrent TACE combined with MTT plus anti-PD-(L)1 immunotherapy. Combination time was defined as TACE and anti-PD-(L)1 treatment at the same time, or MTT and TACE or anti-PD-(L)1 treatment within 60 days before anti-PD-(L)1 treatment, and at least one cycle of anti-PD-(L)1 immunotherapy should be used after TACE. (6) Previous HCC-related treatment history, including only surgical resection. (7) MRI images were available after enhanced MRI examination within 1 month before treatment. Exclusion criteria included: (1) poor image quality and incomplete clinical data; (2) a history of other malignancies besides HCC; (3) tumor lesions that could not be measured or segmented; and (4) other anti-tumor treatments after the start of combined therapy.

[0048] 2 Research Methods

[0049] 2.1 Treatment options

[0050] The aim of this study was to predict the tumor response (including complete response or partial response) of HCC patients to TACE combined with MTT plus anti-PD-(L)1 immunotherapy for 6 months. All patients received standardized conventional TACE treatment, and the entire TACE procedure was performed by physicians from participating centers with at least 10 years of interventional radiology experience.

[0051] In combined immunotherapy, patients took all drugs at standard doses and frequencies according to the guidelines and the supply of PD-(L)1 inhibitors in China. The PD-(L)1 inhibitors used included atezolizumab, carrelizumab, sintilimab, tislelizumab, and toripalimab, and the specific dosing conditions are shown in Table 1 below; for combined molecular targeted drugs, sorafenib, lenvatinib, apatinib, and bevacizumab molecular targeted drugs were administered at standard doses, as shown in Table 2 below.

[0052] Table 1 Dosing criteria for PD-(L)1 inhibitors

[0053]

[0054] Table 2 Dosage criteria for molecular targeted drugs

[0055]

[0056] 2.2 MRI image acquisition

[0057] Patients were followed up every 60 days (±7 days) after starting combination therapy, and tumor response was assessed by contrast-enhanced abdominal MRI / CT and chest CT (CT data were used for patient review and not used to establish a prediction model). Tumor response was assessed according to mRECIST to evaluate the patient's tumor treatment response for 6 months. All follow-up times were at least 6 months.

[0058] MRI image acquisition method: All patients underwent magnetic resonance imaging (MRI) using a 3.0-T scanner (Pioneer, Discovery750, GE Healthcare; Ingenia, Philips Healthcare). MRI sequences included T2-weighted imaging (T2WI), b value of 600s / mm 2 Diffusion weighted imaging (DWI), T1 weighted imaging (T1WI) and dynamic contrast-enhanced imaging were performed. Gadopentetate dimeglumine (Magnevist; Bayer Healthcare, Germany) was injected at a rate of 2 mL / s and a dose of 0.2 mL / kg, and flushed with 20 mL of normal saline. Arterial phase, portal venous phase and delayed phase images were obtained at 20-30 seconds, 60-70 seconds and 180 seconds after injection of contrast agent, respectively. MRI instrument scanning parameters are shown in Table 3.

[0059] Table 3 MRI instrument scanning parameters

[0060]

[0061] 2.3 MRI Image Segmentation and Feature Extraction

[0062] Radcloud software was used for semi-automatic tumor segmentation. After training, one doctor manually drew the regions of interest of all liver tumors along the tumor boundaries on six image sequences: T1WI, T2WI, DWI, arterial phase, portal venous phase, and delayed phase. Another doctor re-segmented 30 randomly selected tumors to test the robustness and reproducibility of the extracted radiological features, as shown in the attached figure. Figure 2 As shown in the attached Figure 2 In the figure, A is T1-weighted sequence; B is T2-weighted sequence; C is diffusion-weighted sequence; D is arterial phase; E is portal phase; F is delayed phase sequence; G is tumor segmentation 3D image. Both doctors were blinded to the clinical and response assessment results of all patients.

[0063] The images were then preprocessed to minimize the effects of different MRI acquisition parameters and scanners. To minimize differences in signal intensity between scanners, the MRI signal intensity was normalized by the mean and standard deviation of all grayscale values ​​in the image. The voxels were then resampled and interpolated using the sitkBSpline interpolator to ensure that all images had the same 1 mm × 1 mm × 1 mm size, thereby minimizing the effects of different MRI imaging acquisition parameters and scanners, and the image grayscale had a binary width of 5.

[0064] The Radcloud platform tool features calculation was used to extract 1688 radiomic features from each modality image, and a total of 10128 features were extracted from the six modalities (http: / / pyradiomics.readthedocs.io / en / latest / features.html). All extracted radiomic features were subdivided into 7 categories, as shown in Table 4. The definition and name of radiomic features conform to the Imaging Biomarker Standardization Initiative (IBSI).

[0065] Table 4 Classification of radiological features

[0066]

[0067]

[0068]

[0069] 2.4 Prediction model development and validation

[0070] The entire process of this study is shown in the attached Figure 3 The entire process includes 9 steps: MRI image acquisition, MRI image preprocessing, semi-automatic tumor segmentation, radiomics feature extraction, radiomics feature selection, radiomics feature display, machine learning model construction, prediction model evaluation, and model clinical application.

[0071] 2.4.1 Feature Selection

[0072] Before the feature selection step, all radiomic features were standardized using the standard scaler function (Python), i.e., removing the mean and dividing by the standard deviation, and then each set of feature values ​​was converted to a mean of 0 and a variance of 1. In order to improve the accuracy of model prediction and reduce the influence of feature redundancy, it was necessary to remove redundant features and select the optimal features. First, the Pearson correlation method (threshold = 0.9) and variance threshold method (variance threshold = 0.08) were used to eliminate features with high correlation or small variance. Then, the Select-K-Best method with P < 0.05 was used to determine the best features associated with the target. Finally, the least absolute shrinkage and selection operator (LASSO) regression method was used for feature screening and dimensionality reduction. The optimal λ is the regularization coefficient in the LASSO method, and the optimal λ was selected by minimizing the mean square error (MSE) using an internal 10-fold cross-validation in the training cohort with a maximum number of iterations of 1500. Subsequently, the radiomic parameters with non-zero coefficients and optimal λ in the LASSO algorithm generated by the entire training cohort were selected. The radiomics score (Rad score) was calculated for the selected features.

[0073] 2.4.2 Prediction model construction

[0074] All prediction models were constructed in the training cohort and validated in two independent validation cohorts to assess the generalizability of the models after they were constructed in the training cohort. The selected features were used as inputs for model construction. Six machine learning models were constructed using machine learning (ML) methods in combination with the above selected features. In the training cohort, the machine learning algorithms used included logistic regression (LR), K-nearest neighbor (KNN), support vector machine (SVM), gradient boosting (GBoost), decision tree, and random forest. During model construction, each classifier was tuned and hyperparameters were optimized to maximize diagnostic performance. The classification results were evaluated by receiver operating characteristic (ROC) curves, and the associated area under the ROC curve (AUC), accuracy, sensitivity, and specificity were obtained. After training, the generalization performance of each model was estimated in the validation cohort.

[0075] 2.4.3 Prediction model evaluation

[0076] The prediction results of the model were presented by confusion matrix. The differences between different ROCs were compared by DeLong test. The optimal cutoff value of the prediction model was determined by ROC by maximizing the Youden index in the training cohort. The net reclassification index (NRI) was used to quantify whether there was a significant difference between the AUC of the new model and the old model compared with the old model. NRI>0 indicated that the predictive ability of the new model was better than that of the old model. The integrated discrimination improvement index (IDI) was used to evaluate the difference in prediction probability between the new and old models, which was calculated based on the prediction probability of the model for each individual. The larger the IDI value, the better the prediction ability of the model. The calibration curve was drawn to evaluate the calibration of the model. The clinical utility of the prediction model was determined by quantifying the net benefit at different thresholds using decision curve analysis.

[0077] 2.4.4 Statistical analysis

[0078] Categorical variables were expressed as counts and percentages and compared using Pearson χ2 analysis, Fisher's exact test, or Mann-Whitney U test. Continuous variables were expressed as mean (± standard deviation) or median (interquartile range (IQR)) and compared using Student's t test or Mann-Whitney U test. Clinical factors with p < 0.05 were identified as clinically independent predictors by univariate and multivariate logistic regression analysis. Scikit-learn package 0.24.2 and Pyradiomics package 3.0.1 were used for feature normalization, feature selection, and model construction. ROC curve analysis was used to evaluate the diagnostic performance of the ML classifier, including the calculation of 95% confidence intervals (CI), specificity, sensitivity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). The clinical model was combined with the best radiomics model to create a joint model. All statistical tests were two-tailed, and P < 0.05 was considered statistically significant. Statistical analyses were performed using Python (version 3.6.8) and R software (version 3.6.1).

[0079] 3 Research results

[0080] 3.1 Patient Baseline Characteristics

[0081] A total of 276 eligible patients were enrolled in this study, 152 entered the training cohort, 60 entered the external validation cohort 1, and 64 entered the external validation cohort 2. There were no significant differences in baseline demographic and clinical factors between the training cohort and the external validation cohort, as shown in Table 5. The tumor treatment response in the training cohort, validation cohort 1 and 2 was 57.9% (88 / 152) (P = 0.360), 56.7% (34 / 60) (P = 0.993), and 50% (32 / 64) (P = 0.360), respectively.

[0082] Table 5 Baseline characteristics of patients in the training cohort and independent validation cohort

[0083]

[0084]

[0085] Tables 6, 7, and 8 below summarize the baseline demographic and clinical characteristics of patients with and without treatment response in the training cohort and external validation cohort after 6 months of combination therapy.

[0086] Table 6 Baseline patient characteristics of patients with and without treatment response in the training cohort

[0087]

[0088]

[0089] Table 7 Baseline patient characteristics of patients with and without treatment response in external validation cohort 1

[0090]

[0091]

[0092] Table 8 Baseline patient characteristics of patients with and without treatment response in external validation cohort 2

[0093]

[0094]

[0095] As can be seen from Tables 6 to 8, in the training cohort, the proportion of patients with AFP ≤ 400 ng / mL who responded to treatment was higher (69.3% vs. 45.3%, P = 0.005) and the NLR was lower (2.41 ± 1.25 vs. 3.67 ± 2.28, P < 0.001) compared with those who did not respond to treatment. In the external validation cohort 1, compared with patients who did not respond to treatment, the proportion of patients with AFP ≤ 400 ng / mL who responded to treatment was higher (67.6% vs. 38.5%, P = 0.047), the NLR was lower (2.43 ± 2.41 vs. 3.96 ± 2.94, P < 0.001), and the proportion of patients without extrahepatic metastasis was lower (64.7% vs. 96.2%, P = 0.009). In the external validation cohort 2, patients with treatment response had a lower NLR compared with those without treatment response (2.78±1.38 vs. 3.87±2.71, P=0.015).

[0096] 3.2 Relationship between clinicopathological features and tumor treatment response

[0097] The results of univariate and multivariate analysis of clinical factors of patients with or without treatment response in the training cohort are shown in Table 9 below. The results show that NLR (P<0.001) and AFP level (>400ng / mL) (P=0.027) are clinical independent predictors of tumor treatment response. Six clinical models of machine learning methods were established based on the two clinical independent predictors. Table 10 summarizes the AUC (95% CI), accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of the clinical models corresponding to the six different machine learning methods. The receiver operating characteristic curve of the clinical model is shown in the attached figure. Figure 4 As shown in Table 10 and the attached Figure 4 As can be seen, the AUC (95% CI) of the clinical logistic regression model was 0.782 (0.682-0.804), 0.695 (0.534-0.732), and 0.679 (0.527-0.719) in the training cohort and external validation cohort, respectively. These results indicate that the clinical model only moderately predicts the tumor treatment response to combination therapy.

[0098] Table 9 Univariate and multivariate analysis results of clinical factors in patients with or without treatment response in the training cohort

[0099]

[0100]

[0101]

[0102] Table 10 Predictive performance of different clinical models in training and external validation cohorts

[0103]

[0104]

[0105] 3.3 Radiomics features predict tumor treatment response

[0106] After radiomics feature selection, the details of the selected features and Rad-score are shown in the attached Figure 5 Finally, 16 radiomic features from 6 modalities were selected to construct the radiomics model, as shown in the attached figure. Figure 5 As shown in C. Figure 5 The correlation heatmap in Figure 2 depicts the correlation coefficient matrix of the 16 radiomic features in the training cohort. The heatmaps of the 16 radiomic features in the three cohorts are shown in the attached figure. Figure 6 As shown. Figure 5 and attached Figure 6As can be seen, 16 features come from different categories. Shape features include: Original_shape_Elongation_T1. First-order statistical features include: Gradient_firstorder_10Percentile_AP, Square_firstorder_Energy_DWI. Gray-level run-length matrix features include: Original_glrlm_ShortRunHighGrayLevelEmphasis_AP, Ibp-3D-k_glrlm_ShortRunEmphasis_DP. Gray-level dependence matrix features include: Original_gldm_DependenceVariance_DWI. The high-order transform wavelet features include: Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_AP, Wavelet-HLL_firstorder_RootMeanSquared_AP, Wavelet-LLL_ngtdm_Strength_AP, Wavelet-LHL_glszm_SmallAreaEmphasis_PP, Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_PP, Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_DP, Wavelet-LLL_gldm_DependenceVariance_DP, Wavelet-LHH_ngtdm_Complexity_T1, Wavelet-LHL_gldm_DependenceEntropy_T2. At the same time, the 16 features come from 6 different modes. The shape feature "Original_shape_Elongation_T1" from the original image T1-weighted sequence reflects the morphological information of the tumor. The first-order feature "Square_firstorder_Energy_DWI" from the DWI sequence reflects the molecular motion of the tumor on the DWI sequence. The feature "Original_glrlm_ShortRunHighGrayLevelEmphasis_AP" from the arterial phase may be related to the tumor enhancement in the arterial phase.The feature "Wavelet-LHL_glszm_SmallAreaEmphasis_PP" from the arterial phase reflects the uniformity of the image texture. The larger its feature value, the more uneven the image texture. The feature "Wavelet-LLL_gldm_DependenceVariance_DP" from the delayed phase reflects the grayscale discreteness of the tumor area, which may be related to tumor heterogeneity. In short, the more malignant HCC is, the greater the difference in its cell composition, the more obvious the heterogeneity in and around the tumor, and the less necrosis of tumor tissue after combined treatment, but still significantly enhanced, which is manifested in the image as increased image grayscale inhomogeneity, increased image complexity, and more uneven texture.

[0107] Attached Figure 7 Center AC shows the radiomic Rad scores of patients in the training cohort and the external validation cohort. The relationship between Rad scores and tumor response in the three cohorts is shown in the attached figure. Figure 7 The results showed that Rad-score was significantly correlated with tumor response (P<0.001).

[0108] Based on the 16 radiomic features screened above, 6 machine learning models were used to establish the radiomics model. The AUCs of the 6 machine learning models are shown in the attached figure. Figure 8 The accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the six machine learning models are shown in Table 11.

[0109] Table 11 Predictive performance of different radiomics models in training and external validation cohorts

[0110]

[0111] Through model comparison, the logistic regression model with the best predictive efficacy was finally selected. In the training cohort, the optimal cutoff value based on the maximum Youden index was 0.535, which effectively predicted patients with and without treatment response in the three cohorts.

[0112] 3.4 Clinical-radiomics joint model

[0113] In this study, radiomic features were combined with clinical factors to construct a clinical-radiomics joint model. The AUCs of the six joint models are shown in the attached figure. Figure 8DF; the accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the joint model are shown in Table 12. As can be seen from Table 12, among the six machine learning joint models, the logistic regression model has the best predictive efficiency, and its AUC (95% CI) is 0.956 (0.859-0.942), 0.895 (0.738-0.903), and 0.892 (0.667-0.828) in the training and external validation cohorts, respectively. The nomogram based on the joint model of logistic regression is shown in the attached figure. Fig. 9 The nomogram based on clinical factors combined with radiomic features can predict the treatment response of each patient. By combining the patient's clinical factors and radiomic scores, the total score of each patient can be obtained, thereby predicting the treatment efficacy.

[0114] Table 12 Prediction performance of the combined model in training and external validation cohorts

[0115]

[0116] Attached Fig.10 Figure 2 shows the radar charts of the clinical model, radiomics model and joint model and the confusion matrix of the joint model under the logistic regression model. AC is the radar chart of the clinical model, radiomics model and joint model, and DF is the confusion matrix of the joint model in the three queues. Fig.10 It can be seen that the joint model has the best prediction performance, and the joint model based on logistic regression can effectively predict the patient's treatment response.

[0117] 3.5 Model Evaluation

[0118] The calibration curve and decision curve of the joint model are shown in the attached figure. Fig.11 As shown, AC is the calibration curve of the joint model in the three cohorts. DF is the decision curve of the joint model in the three cohorts. The results showed that there was good consistency between the results predicted by the joint model and the actual clinical results. The decision curve showed that the joint prediction model could increase the net benefit compared with assuming that all patients had or did not have treatment response in the training and external validation cohorts. These results show that the radiomic features of pre-treatment multimodal MRI combined with clinical factors can predict tumor treatment response to TACE combined with MTT plus anti-PD-(L)1 immunotherapy, with good discrimination, calibration and clinical utility.

[0119] 3.6 Incremental predictive value of radiomics features over clinical factors

[0120] The incremental value results of the radiomics model relative to the clinical model are shown in Table 13. Compared with the clinical model, the AUC of the clinical-radiomics combined model in the three cohorts was significantly improved, namely (0.956 vs. 0.782, P < 0.001), (0.895 vs. 0.695, P < 0.001), and (0.892 vs. 0.679, P < 0.001), respectively. According to the net reclassification improvement index (NRI) and the integrated discrimination improvement index (IDI), the predictive performance of the clinical-radiomics combined model was significantly improved compared with the clinical model in the training cohort (NRI = 1.469, P < 0.001; IDI = 49.2%, P < 0.001), external validation cohort 1 (NRI = 0.950, P < 0.001; IDI = 28.8%, P < 0.001), and external validation cohort 2 (NRI = 1.063, P < 0.001; IDI = 31.5%, P < 0.001). These results suggest that radiomics features can provide incremental value to clinical factors in predicting tumor treatment response.

[0121] Table 13 Incremental value of radiomics models relative to clinical models

[0122]

[0123] In summary, the radiomic features extracted from pre-treatment multimodal enhanced MRI can effectively predict the personalized tumor treatment response of patients with advanced unresectable HCC to TACE combined with MTT plus anti-immunotherapy. The clinical-radiomic combined model has the best predictive efficacy, and the radiomic features provide incremental predictive value for clinical factors.

[0124] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features, characterized by: The nomogram of the prediction model includes: The first line is a score scale, ranging from 0 to 100; The second row shows the alpha-fetoprotein AFP values, including ≤400 and >45; The third row is the neutrophil-to-lymphocyte ratio (NLR), ranging from 0 to 11; The fourth line is the radiomics score Rad-score, ranging from -1 to 0.6; The fifth line is the total score, which is the sum of the numbers in the second to fourth lines, ranging from 0 to 140; The sixth line is the linear prediction value, ranging from -12 to 8; The seventh line is the prediction value of treatment effect, ranging from 0.1 to 0.

95.

2. The method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features according to claim 1, characterized in that: The following steps are involved: S1: Obtain contrast-enhanced MRI images of the abdomen of patients with liver cancer before starting combined therapy; S2: Segment the contrast-enhanced abdominal MRI images of patients to obtain radiomic features; S3: Selection of clinically independent predictors and radiomic features for building a prediction model; S4: Use different machine learning methods to build clinical models and radiomics models respectively; S5: Perform model evaluation on the model constructed in step S4 to determine the best clinical model and radiomics model; S6: combining the best clinical model and radiomics model determined in step S5 to obtain a joint model, i.e., the final prediction model; S7: Conduct external validation of the constructed joint model.

3. The method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features according to claim 2, characterized in that: The patient's abdominal MRI contrast-enhanced images described in step S1 include six modality images: T1WI, T2WI, DWI, arterial phase, portal venous phase, and delayed phase.

4. The method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features according to claim 3, characterized in that: The clinical independent predictive factors described in step S3 include the neutrophil-to-lymphocyte ratio (NLR) and alpha-fetoprotein (AFP) value.

5. The method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features according to claim 4, characterized in that: The radiomic features described in step S3 include Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_AP, Original_glrlm ShortRunHighGrayLevelEmphasis_AP, Wavelet-LLL_glrlm_ShortRunHighGrayLevelEmphasis_PP, Wavelet-LLL_ngtdm_Strength_AP, Gradient_firstorder_10Percentile_AP, Wavelet-HLL_firstorder_RootMeanSquared_AP, Wavelet-LL_gldm_DependenceVariance_DP, Exponential_gldm_DependenceVariance_DP, Original_g ldm_DependenceVariance_DWI, Wavelet-LHL_gldm_DependenceEntropy_T2, lbp-3D-k_glrlm_ShortRunEmphasis_DP, Wavelet-LL_glrlm_ShortRunHighGrayLevelE mphasis_DP, square_firstorder_Energy_DWI, Wavelet-LHL_glszm_SmallAreaEmphasis_PP, Wavelet-LHH_ngtdm_Complexity_T1 and Original_shape_Elongation_T1.

6. The method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features according to claim 5, characterized in that: The machine learning methods described in step S4 include logistic regression, K-nearest neighbor, support vector machine, gradient boosting, decision tree and random forest.

7. The method for constructing a prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features according to claim 6, characterized in that: The best clinical model and radiomics model described in step S5 are the clinical model and radiomics model of logistic regression.

8. Use of the prediction model for predicting the effect of combined treatment of liver cancer based on multimodal MRI radiomics features as claimed in claim 1 in the selection of treatment plans before treatment of liver cancer patients.

Images