An imaging genomics and clinical nonlinear fusion prediction method

Abstract

The application discloses a kind of imageomics and clinical nonlinear fusion prediction methods, it is related to image processing technical field, by acquiring the prostate double parameter magnetic resonance T2WI and DWI image data of having been collected, based on the initialization information provided by user is obtained with semi-automatic segmentation and extracts high-dimensional imageomics features, then in turn executes feature correlation rejection, LASSO sparseness compression and Bootstrap resampling stability score three-level classification screening to obtain stable target feature subset and calculate imageomics score, and based on existing clinical feature data calculation clinical score, by including the product interaction of imageomics score and clinical score Nonlinear weighted fusion function calculation joint prediction score, finally output auxiliary prediction data and visual result.The application improves the segmentation efficiency of region of interest, improves the accuracy, stability and generalization performance of Ki-67 expression level prediction data.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for predicting nonlinear fusion of radiomics and clinical data. Background Technology

[0002] Prostate cancer is one of the most common malignant tumors of the male urinary system, and accurate assessment of its biological invasiveness is crucial for clinical decision-making. Ki-67, a nucleoprotein associated with cell proliferation, reflects the proliferative activity of tumor cells and is closely related to the differentiation, invasiveness, and prognosis of prostate cancer. Currently, obtaining Ki-67 expression levels mainly relies on immunohistochemical staining of biopsy or postoperative pathological specimens, which is an invasive procedure with a risk of sampling bias and makes it difficult to comprehensively assess the overall biological behavior of the tumor before treatment. In recent years, radiomics technologies based on magnetic resonance imaging (MRI) have been gradually applied to tasks such as the detection and invasiveness assessment of prostate cancer. Among these, dual-parameter MRI, due to its omission of dynamic contrast-enhanced sequences, offers advantages such as shorter scan time, lower cost, and no need for contrast agent injection, making it an important image source for prostate radiomics research. Existing studies have attempted to extract radiomics features from bp-MRI images and combine them with clinical parameters to construct predictive models for Ki-67 expression levels. However, existing approaches still have the following shortcomings: In the feature selection stage, most methods only use a single correlation coefficient threshold to remove redundant features, which easily retains statistical noise with limited samples, thus limiting the model's generalization ability; in the fusion stage of radiomics features and clinical features, a simple linear weighted combination method is usually used, which is difficult to fully capture the nonlinear synergistic relationship that may exist between the two types of information; in the acquisition of regions of interest, it generally relies on physicians to manually delineate lesions layer by layer, which is inefficient and has poor inter-observer consistency. In addition, there is still a lack of a comprehensive data processing scheme that can simultaneously take into account the stability of feature selection, nonlinear fusion of multi-source information, and segmentation efficiency. Summary of the Invention

[0003] To simultaneously ensure the stability of feature selection, nonlinear fusion of multi-source information, and segmentation efficiency, this invention proposes a nonlinear fusion prediction method combining radiomics and clinical data, comprising the following steps: S1: Obtain the acquired prostate dual-parameter magnetic resonance image data of the object to be processed, the data including T2-weighted imaging image data and diffusion-weighted imaging image data; S2: Based on the segmentation initialization information provided by the user, the T2-weighted imaging image data and diffusion-weighted imaging image data are processed using a semi-automatic segmentation method to obtain the region of interest; S3: By extracting radiomics features from the region of interest, a high-dimensional initial set of radiomics features is obtained; S4: Obtain a subset of target image omics features by performing hierarchical screening on the initial image omics feature set. The hierarchical screening includes feature correlation elimination, LASSO sparsity compression and Bootstrap resampling stability score screening performed sequentially. S5: Based on a subset of target image omics features, an image omics score is calculated using a pre-defined scoring model. S6: Obtain the existing clinical feature data of the object to be processed, and calculate the clinical score based on the clinical feature data through a preset clinical scoring model. S7: Based on a nonlinear weighted fusion function that includes a product interaction term of radiomics scores and clinical scores, the radiomics scores and clinical scores are fused to calculate a joint predictive score. S8: Output the joint prediction score and its visualization results as auxiliary prediction data related to the expression level of Ki-67 in the object to be processed.

[0004] This invention improves the efficiency of region of interest acquisition through semi-automatic segmentation, enhances the cross-sample stability and model generalization ability of radiomics features through a three-level hierarchical screening of feature correlation elimination, LASSO sparsity compression, and Bootstrap resampling stability score, and fully captures the synergistic effect between multi-source information through a nonlinear weighted fusion function that includes the interaction term of the product of radiomics score and clinical score. The three elements work synergistically in their respective dimensions to achieve a comprehensive improvement in the accuracy, stability, and processing efficiency of Ki-67 expression level-assisted prediction data.

[0005] Furthermore, in step S2, the user-provided segmentation initialization information is the seed point coordinate data of the lesion center region, and the semi-automatic segmentation method is the region growing algorithm.

[0006] Furthermore, the region growth algorithm specifically involves: according to a preset neighborhood connectivity criterion and a gray-level difference threshold, including adjacent pixels whose gray-level difference with the seed point is less than the gray-level difference threshold into the growth region until a preset termination condition is met.

[0007] Furthermore, in step S4, feature correlation elimination specifically involves: calculating the correlation coefficient between any two features in the initial radiomics feature set; when the absolute value of the correlation coefficient exceeds a preset threshold, eliminating features with weak correlation to the prediction target to obtain a subset of candidate radiomics features.

[0008] Furthermore, in step S4, LASSO sparsity compression specifically involves: constructing a LASSO regression model for a subset of candidate image omics features, using cross-validation to determine the optimal penalty coefficient, and removing features whose regression coefficients are compressed to zero.

[0009] Furthermore, in step S4, the Bootstrap resampling stability score selection is specifically as follows: the training data is resampled multiple times with replacement, and LASSO sparsity compression is repeatedly performed after each resampling. The frequency ratio of each candidate feature being selected is counted as the stability score, and only features with stability scores not lower than a preset threshold are retained as the target image omics feature subset.

[0010] Furthermore, in step S6, the clinical characteristic data includes at least one of the following: PI-RADS score, Gleason score, and prostate-specific antigen level.

[0011] Furthermore, in step S7, the nonlinear weighted fusion function further includes a nonlinear transformation term based on radiomics scoring, and the fusion weight coefficients of the product interaction term and the nonlinear transformation term are determined by optimization on the training dataset using a combination of multivariate logistic regression and grid search.

[0012] Furthermore, the nonlinear transformation function corresponding to the nonlinear transformation term is the Sigmoid function, which is used to map the radiomics score to a preset numerical range, thereby enhancing the ability of the joint prediction score to represent the nonlinear information in the radiomics score.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The radiomics and clinical nonlinear fusion prediction method proposed in this invention, in the radiomics feature screening stage, through the three-level screening mechanism consisting of correlation elimination, LASSO sparsity compression and Bootstrap resampling stability score, further screens out the target feature subset with high cross-sample stability on the basis of eliminating redundant features, effectively reducing the risk of overfitting and improving the generalization performance of the model. (2) In the information fusion stage, by introducing the product interaction term of radiomics score and clinical score, the fusion function can represent the synergistic effect between the two information sources. On this basis, a nonlinear transformation term is further introduced to enhance the ability to represent nonlinear information in radiomics score. Compared with the single linear weighted fusion method, it can more fully explore the complementary relationship between multiple sources of information. (3) In the acquisition of region of interest, a semi-automatic segmentation method based on seed point growth is adopted, which simplifies the doctor's operation to two steps: single point selection and boundary confirmation. This effectively shortens the processing time of a single case and reduces the segmentation difference between different operators. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the steps of a nonlinear fusion prediction method combining radiomics and clinical practice. Figure 2 A nodal plot constructed based on joint prediction scores; Figure 3 Feature curves for the training and validation datasets; Figure 4 Calibrate the curves for the training dataset; Figure 5 To validate the calibration curves of the dataset; Figure 6 This is a graph showing the decision curve analysis of the joint prediction model. Detailed Implementation

[0015] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments.

[0016] This invention provides a radiomics and clinical nonlinear fusion prediction method that directly processes acquired magnetic resonance imaging (MRI) data. Its output is an auxiliary data indicator reflecting the probability of Ki-67 expression levels. It does not involve performing MRI scans on a living human body, nor does it directly derive a disease diagnosis or treatment plan. In practical applications, it serves only as an auxiliary data processing tool for medical professionals to make clinical decisions, and its use must be combined with the professional judgment of a licensed physician. Figure 1 As shown, the prediction method mainly includes the following steps: S1: Obtain the acquired prostate dual-parameter magnetic resonance image data of the object to be processed, the data including T2-weighted imaging image data and diffusion-weighted imaging image data; S2: Based on the segmentation initialization information provided by the user, the T2-weighted imaging image data and diffusion-weighted imaging image data are processed using a semi-automatic segmentation method to obtain the region of interest; S3: By extracting radiomics features from the region of interest, a high-dimensional initial set of radiomics features is obtained; S4: Obtain a subset of target image omics features by performing hierarchical screening on the initial image omics feature set. The hierarchical screening includes feature correlation elimination, LASSO sparsity compression and Bootstrap resampling stability score screening performed sequentially. S5: Based on a subset of target image omics features, an image omics score is calculated using a pre-defined scoring model. S6: Obtain the existing clinical feature data of the object to be processed, and calculate the clinical score based on the clinical feature data through a preset clinical scoring model. S7: Based on a nonlinear weighted fusion function that includes a product interaction term of radiomics scores and clinical scores, the radiomics scores and clinical scores are fused to calculate a joint predictive score. S8: Output the joint prediction score and its visualization results as auxiliary prediction data related to the expression level of Ki-67 in the object to be processed.

[0017] Overall, the entire prediction process includes key steps such as image data acquisition, region of interest segmentation, radiomics feature extraction, hierarchical feature screening, radiomics score calculation, clinical score calculation, joint prediction score fusion, and result output.

[0018] In the image data acquisition stage, the first step is to acquire the pre-existing dual-parameter magnetic resonance imaging (MRI) data of the prostate gland. This dual-parameter MRI data includes T2-weighted imaging data and diffusion-weighted imaging data. In practical applications, the dual-parameter MRI images of the prostate gland can be acquired using conventional MRI scanning equipment. However, the method of this invention does not involve in vivo scanning; instead, it directly utilizes the existing image data for subsequent processing. T2-weighted imaging and diffusion-weighted imaging sequences are the two most commonly used sequences in prostate dual-parameter MRI. T2-weighted imaging can clearly display the anatomical structure of the prostate and the morphological characteristics of lesions, while diffusion-weighted imaging can reflect the degree of diffusion restriction of water molecules within the lesions. The two complement each other, providing multi-dimensional image information for subsequent radiomics analysis. The acquired image data is typically stored in DICOM format, which includes the image pixel matrix and corresponding spatial location information. In practice, the image acquisition module can read the image data from the storage medium and convert it into a data format suitable for subsequent processing steps, such as converting DICOM format to NIfTI format or other formats that facilitate image processing.

[0019] After acquiring T2-weighted and diffusion-weighted imaging data, the lesion region in the image needs to be segmented to determine the region of interest (ROI) upon which subsequent feature extraction depends. In one specific embodiment of the present invention, the ROI is determined using a seed point-based region growing algorithm to achieve semi-automatic segmentation. Specifically, a radiologist with experience in interpreting prostate MRI images selects a pixel as a seed point by clicking on the inner side of the lesion region in the T2-weighted and diffusion-weighted imaging images displayed on the display device. The two-dimensional image coordinates of the seed point and its layer position information together constitute the seed point coordinate data. Compared with the fully manual segmentation method in the prior art, which requires radiologists to completely delineate the boundaries of the lesion layer by layer, the semi-automatic segmentation method adopted in the present invention simplifies the radiologist's operation to selecting a single seed point, and the subsequent growing process is automatically completed by the computer, reducing operational complexity and processing time.

[0020] After the seed point is selected, the region growing algorithm uses this seed point as the starting point and, according to preset neighborhood connectivity and gray-level similarity criteria, gradually incorporates neighboring pixels whose gray-level difference with the seed point is less than a preset gray-level difference threshold into the growing region. The neighborhood connectivity criterion can be four-neighbor connectivity or eight-neighbor connectivity; in this embodiment, eight-neighbor connectivity is preferred, meaning that the seed point is the center, and its eight surrounding neighboring pixels are examined. The gray-level similarity criterion is based on whether the absolute value of the difference between the gray-level value of a pixel and the current seed point is less than a preset threshold. When the gray-level difference of a neighboring pixel meets the threshold condition, that pixel is incorporated into the growing region and becomes a new starting point for growth, continuing to expand outwards. This process is iterated until no new pixels meet the inclusion condition. When segmenting different image sequences, the specific value of the gray-level difference threshold can be adaptively adjusted based on factors such as the model of the image acquisition device, scanning parameters, and the contrast-to-noise ratio of the image itself. A reasonable grayscale difference threshold can ensure that the segmented region of interest can cover the solid part of the lesion well, while avoiding excessive extension into the surrounding normal tissue area.

[0021] After obtaining the region of interest (ROI), radiomics feature extraction is performed on the ROI to obtain an initial radiomics feature set. Radiomics feature extraction refers to the quantitative extraction of a large number of descriptive features from medical images, transforming the visual information in the images into numerical data that can be analyzed and modeled by computers. In this embodiment, feature extraction is performed on the ROI of two sequences, T2-weighted imaging and diffusion-weighted imaging, respectively. The extracted feature types include first-order statistical features, shape features, gray-level co-occurrence matrix features, gray-level region size matrix features, gray-level run-length matrix features, and derived texture features after various image preprocessing transformations. The first-order statistical features describe the distribution of pixel grayscale values ​​within the region of interest, including statistical quantities such as mean, variance, kurtosis, skewness, energy, and entropy. Shape features describe the geometric morphology of the region of interest, including three-dimensional morphological parameters such as surface area, volume, sphericity, and compactness. The gray-level co-occurrence matrix feature describes the image's texture roughness, directionality, and contrast by calculating the probability of simultaneous occurrence of grayscale values ​​in pixel pairs with specific spatial relationships. The gray-level region size matrix feature and gray-level run length matrix feature characterize the image's texture uniformity and structure from the perspectives of region size and run length, respectively. To further enrich the feature information, the original image undergoes preprocessing operations such as wavelet transform, Laplacian Gaussian filtering, and local binary mode transform to generate various derived images, and the aforementioned features are extracted from these derived images. Taking this embodiment as an example, after performing the aforementioned feature extraction on T2-weighted imaging and diffusion-weighted imaging data of 114 subjects, 1923 initial radiomics features were obtained for each subject. It is worth noting that this number of features is only an illustrative example. In practical applications, the number of initial radiomics features will vary depending on the image acquisition parameters, feature extraction algorithm configuration, and preprocessing methods. However, they are usually high-dimensional feature sets, meaning that the feature dimension is much larger than the number of samples. This is also a common "curse of dimensionality" problem in radiomics data processing.

[0022] Since the initial radiomics feature set usually contains a large number of redundant features and noisy features that may be irrelevant to the prediction target, directly using them for modeling can lead to overfitting and decreased generalization ability. Therefore, it is necessary to screen the initial feature set to select a subset of features that are highly correlated with the prediction target and have high cross-sample stability. In this invention, the feature screening adopts a hierarchical screening strategy, which includes three steps: feature correlation elimination, LASSO sparsity compression, and Bootstrap resampling stability scoring. Through the progressive screening of these three steps, redundant, collinear, and occasionally significant features are gradually eliminated, and finally a stable subset of target radiomics features is obtained.

[0023] In the feature correlation elimination stage, the Pearson correlation coefficient between any two features in the initial radiomics feature set is first calculated. When the absolute value of the correlation coefficient between two features exceeds a preset threshold, it is determined that the two features have a high linear correlation and information redundancy. At this time, the association strength between the two features and the prediction target is calculated separately, and features with weaker association strength are eliminated. In this invention, the prediction target is the grouping label of Ki-67 expression level, which is derived from existing pathological immunohistochemical assessment results. The association strength can be measured by the univariate logistic regression coefficient between the feature and the label, mutual information, or other indicators that can quantify the correlation between the feature and the classification target. After this process, 1909 of the initial 1923 features were eliminated, leaving only 14 candidate radiomics features. It should be noted that the correlation coefficient threshold can be flexibly adjusted according to the actual number of data features and sample size, generally ranging from 0.2 to 0.5. The lower the threshold, the stricter the feature retention and the greater the elimination force.

[0024] After relevance removal, the remaining candidate radiomics features are subjected to LASSO sparsity compression. LASSO regression is a linear regression method based on L1 norm regularization. Its basic idea is to introduce the sum of the absolute values ​​of all regression coefficients as a penalty term into the regression objective function, thereby compressing some unimportant regression coefficients to zero during parameter estimation, achieving feature selection and model sparsity. In this embodiment, a LASSO regression model is constructed based on the aforementioned 14 candidate radiomics features. The objective function of the model is the sum of squared residuals plus the penalty coefficient multiplied by the sum of the absolute values ​​of the regression coefficients. By adjusting the size of the penalty coefficient, the sparsity of feature selection can be controlled: the larger the penalty coefficient, the more regression coefficients are compressed to zero, and the fewer features are retained; the smaller the penalty coefficient, the more features are retained. The formula is expressed as follows: , In the formula, The total number of samples in the training dataset, This represents the total number of candidate radiomics features. The pathological grouping label for the Ki-67 expression level corresponding to the i-th sample is... Let be the candidate image omics feature vector for the i-th sample. Let be the vector of regression parameters to be estimated. Let be the regression coefficient corresponding to the j-th candidate radiomics feature. Let be the absolute value of the regression coefficient of the j-th candidate radiomics feature (as part of the L1 norm regularization penalty term). The regularization penalty coefficient is preset or optimized through cross-validation. The optimal penalty coefficient is determined using a ten-fold cross-validation method, where the training dataset is randomly divided into ten equal parts. One part is used as the validation set, and the remaining nine parts are used as the training set for cross-validation. The mean squared error of the cross-validation model under different penalty coefficients is calculated, and the penalty coefficient value corresponding to the minimum mean squared error is selected, or the penalty coefficient value corresponding to the simplest model (i.e., the fewest features) within one standard error range of the minimum mean squared error is selected. In this embodiment, the latter strategy is used to determine the optimal penalty coefficient to ensure a good balance between predictive performance and simplicity. Through LASSO sparsity compression, some of the 14 candidate features are removed, further reducing the number of features.

[0025] However, LASSO compression results can be affected by random fluctuations in data when based on a single training dataset, leading to insufficient stability of the selected features across different training sample subsets. To overcome this problem, this invention further introduces a Bootstrap resampling stability score for screening, building upon relevance removal and LASSO compression. Bootstrap resampling is a statistical inference method based on sampling with replacement. It generates multiple Bootstrap samples of the same size as the original dataset by performing multiple samplings with replacement on the original training dataset. In each Bootstrap sample, approximately 63.2% of the samples come from the original dataset, with the remainder being duplicate samples.

[0026] In this embodiment, the training dataset is resampled 1000 times with replacement. After each resampling, the aforementioned LASSO sparsity compression process is repeated on the new samples, and the selection of each candidate feature in the LASSO model is recorded. For each candidate feature, the number of times it is selected in all 1000 resampling iterations is counted, and the ratio of this number to the total number of resampling iterations is defined as the stability score of the feature. A higher stability score indicates a higher selection frequency of the feature across different data subsets and better cross-sample repeatability. Subsequently, only features with a stability score not lower than a preset threshold are retained as the final target imageomics feature subset. After this three-level screening, the final target imageomics feature subset obtained in this embodiment contains 8 features. These eight features are derived from different sequences and preprocessed images, specifically including correlation features based on the three-dimensional gray-level co-occurrence matrix after Laplacian Gaussian filtering, large-area low-gray-level emphasis features based on the three-dimensional gray-level region size matrix after Laplacian Gaussian filtering, first-order statistical root mean square, maximum, and skewness features based on different sub-bands after wavelet transform, and clustering shadow features based on the gray-level co-occurrence matrix after wavelet transform. These features reflect the texture complexity and gray-level distribution characteristics of lesions from different perspectives and are highly correlated with the proliferative activity of tumor cells.

[0027] After obtaining the target radiomics feature subset, a radiomics score is calculated based on this feature subset. In this embodiment, a logistic regression model is used to construct the radiomics score function. Eight features from the target radiomics feature subset are used as input variables, and the grouping label of Ki-67 expression level is used as the dependent variable. The logistic regression model is fitted on the training dataset to obtain the regression coefficients and intercept term for each feature. The radiomics score is the sum of the intercept term and the product of each feature value and its corresponding regression coefficient. The fitting of the logistic regression model can be performed using the maximum likelihood estimation method, solving for the maximum likelihood estimate of the regression coefficients through an iterative optimization algorithm. This radiomics score is essentially a continuous numerical value; its level reflects the probability of predicted high Ki-67 expression based on the radiomics features. A higher score indicates a greater probability of predicted high Ki-67 expression.

[0028] On the other hand, this invention also utilizes existing clinical characteristic data of the subject to be processed to calculate a clinical score. Clinical characteristic data includes, but is not limited to, the Prostate Imaging Reporting and Data System (PIDS) score, the Gleason score, and prostate-specific antigen (PSA) levels. The PIDS score is a grading system based on multi-parameter magnetic resonance imaging for standardized assessment of prostate lesions; a higher score indicates a greater likelihood of clinically significant prostate cancer. The Gleason score is a prostate cancer grading system based on pathological histology, reflecting the degree of pathological differentiation of the tumor. PSA is a commonly used serological marker for prostate cancer, and elevated levels are associated with the occurrence and progression of prostate cancer. In practical applications, other clinical factors such as age and lesion location may also be considered; however, according to the study in this embodiment, the association between age and lesion location and Ki-67 expression levels did not reach statistical significance. Therefore, the PIDS score, Gleason score, and PSA level are primarily used as the basis for calculating the clinical score. Specifically, firstly, univariate logistic regression analysis was performed on each candidate clinical feature to screen out features that were statistically significantly associated with Ki-67 expression levels. Then, these features were incorporated into a multivariate logistic regression model, and the model was fitted on the training dataset to obtain the regression coefficients and intercepts for each clinical feature. The clinical score is the sum of the intercept and the value of each clinical feature multiplied by its corresponding regression coefficient. In this embodiment, the associations between the Prostate Imaging Reporting and Data System (PIRS) score, Gleason score, and prostate-specific antigen (PSA) level and Ki-67 expression levels were all statistically significant; therefore, these three items were included in the clinical scoring model.

[0029] After obtaining the radiomics score and clinical score separately, the two are fused to calculate a joint predictive score. Unlike the simple linear weighted summation method in existing technologies, this invention employs a nonlinear weighted fusion function that includes a product interaction term and a nonlinear transformation term. Specifically, the joint predictive score is calculated as follows: radiomics score multiplied by a first weight coefficient, clinical score multiplied by a second weight coefficient, the product of the radiomics score and clinical score multiplied by a third weight coefficient, and the nonlinear transformation value of the radiomics score multiplied by a fourth weight coefficient. The product interaction term of the radiomics score and clinical score is used to capture the synergistic effect between them; that is, when both the radiomics score and clinical score are at a high level, their interaction may have an impact on the prediction result greater than the sum of their individual contributions. The nonlinear transformation term is used to capture nonlinear information that may exist in the radiomics score, further enhancing the fitting ability of the joint predictive score to the original data. In this embodiment, the nonlinear transformation function uses the Sigmoid function, which can map the radiomics score to a numerical range between 0 and 1, giving the radiomics score better numerical stability and interpretability after transformation. The first to fourth weight coefficients mentioned above are collectively referred to as fusion weight coefficients. They are determined by optimizing a grid search strategy on the training dataset based on multivariate logistic regression. Grid search is a hyperparameter optimization method that evaluates the model performance metrics under each preset candidate weight coefficient combination, selecting the combination with the best performance as the final fusion weight coefficients. In this embodiment, the AUC value is used as the performance evaluation metric. The performance of each weight coefficient combination is evaluated on the training dataset using five-fold cross-validation, ultimately determining a set of weight coefficients that maximizes the cross-validation AUC value.

[0030] Based on the joint prediction score mentioned above, a nomogram can be constructed to visually present the prediction model. A nomogram is a commonly used tool for visualizing statistical prediction models. It quantifies the contribution of each variable in the regression model in the form of integrals, obtains the total score by summing the integrals of each variable, and then maps the total score to the predicted probability. Figure 2This paper presents a nomogram constructed based on the joint prediction model of this embodiment. In the nomogram, each variable (including GS (Gleason score), radscore (radiomics score), PSA (prostate-specific antigen), PIRADS score, Total Points (total score scale), and ki67 (predicted probability)) is assigned a corresponding integral scale based on its regression coefficient. Users can find the corresponding integral value (Points, score scale) based on the value of the object to be processed on each variable, sum all integral values ​​to obtain the total score, and then read the corresponding Ki-67 high expression predicted probability based on the mapping relationship between the total score and the predicted probability. The nomogram provides clinicians with a simple and intuitive tool, enabling the convenient application of complex mathematical model results in clinical practice.

[0031] To verify the technical effectiveness of the method proposed in this invention, this embodiment conducted a detailed efficacy evaluation based on prostate cancer case data from an acquired retrospective imaging database. The data came from dual-parameter magnetic resonance imaging data of 114 subjects diagnosed with prostate cancer by pathological examination, randomly divided into a training dataset (80 cases) and a validation dataset (34 cases) at a ratio of 7:3. Among them, 96 subjects (84.2%) showed low Ki-67 expression (positive rate less than 25%), and 18 subjects (15.8%) showed high Ki-67 expression (positive rate greater than or equal to 25%). 28 subjects had a Gleason score less than 7, 57 subjects had a score between 7 and 8, and 29 subjects had a score greater than 8; 4 subjects had a Prostate Imaging Reporting and Data System (PIDS) score of 1, 11 subjects had a score of 2, 25 subjects had a score of 3, and a total of 74 subjects had scores of 4 and 5.

[0032] To comprehensively evaluate the predictive efficacy of the nonlinear fusion joint prediction model of this invention, a clinical scoring model and a radiomics scoring model were constructed as controls. The clinical scoring model was based solely on clinical feature data (Prostate Imaging Reporting and Data System score, Gleason score, and prostate-specific antigen level) to build a logistic regression model, excluding radiomics features. The radiomics scoring model was based solely on the radiomics scores obtained in step S5 to build a logistic regression model, excluding clinical features. The three models were evaluated on both the training and validation datasets. Evaluation metrics included the area under the receiver operating characteristic (ROC) curve and its 95% confidence interval, accuracy, sensitivity, and specificity. The ROC curve was plotted with the false positive rate on the x-axis and the true positive rate on the y-axis; a larger area under the curve indicated stronger discriminative ability. Accuracy reflected the proportion of correct predictions; sensitivity reflected the model's ability to correctly identify positive samples; and specificity reflected the model's ability to correctly identify negative samples. Figure 3The chart shows a comparison of the receiver operating characteristic (ROC) curves of the three models on the training and validation datasets. The vertical axis represents sensitivity, and the horizontal axis represents the false positive rate (1 - specificity). The left side shows the ROC curve for the training dataset, and the right side shows the ROC curve for the validation dataset. Table 1 lists the specific evaluation values ​​of the three models on the training and validation datasets.

[0033] Table 1: Comparison of the predictive power of the three models for Ki-67 expression levels The above results demonstrate that, on both the training and validation datasets, the joint prediction model of this invention outperforms the standalone clinical scoring model and the standalone radiomics scoring model in terms of area under the curve, accuracy, sensitivity, and specificity. In terms of the improvement in area under the curve, on the training dataset, the joint model improves accuracy by approximately 8.0% compared to the standalone radiomics scoring model and by approximately 22.1% compared to the standalone clinical scoring model; on the validation dataset, the joint model improves accuracy by approximately 8.2% compared to the standalone radiomics scoring model and by approximately 13.6% compared to the standalone clinical scoring model. These quantitative data fully validate the effectiveness of the hierarchical feature selection mechanism and the nonlinear weighted fusion mechanism containing product interaction terms proposed in this invention. The hierarchical feature selection mechanism effectively eliminates unstable features with low selection frequency in repeated sampling through the Bootstrap resampling stability score, making the radiomics features that finally enter the model have stronger cross-sample repeatability, thereby reducing the performance degradation of the model on the validation dataset; the nonlinear weighted fusion mechanism captures the synergistic effect between radiomics scores and clinical scores through the product interaction term. Compared with the simple linear weighted summation method, it can make fuller use of the complementary relationship between multi-source information, thereby improving prediction accuracy.

[0034] Furthermore, this embodiment also uses a calibration curve to evaluate the consistency between the predicted probabilities output by the joint prediction model and known pathological reference labels. The calibration curve plots the predicted probability on the x-axis and the actual observed positive event rate on the y-axis, showing the relationship between the predicted probability and the observed probability. Ideally, the calibration curve should be close to the diagonal, meaning that the predicted probability and the observed probability are essentially consistent. Figure 4 The calibration curves for the training dataset are shown. Figure 5 The calibration curves on the validation dataset are shown. Figure 4 and Figure 5As can be seen, on both the training and validation datasets, the calibration curves between the predicted probability (horizontal axis Predicted Pr (ki67=1), Apparent line in the figure) and the observed probability (vertical axis Actual Probability, Bias-corrected line in the figure) are close to the diagonal line (Ideal line in the figure), indicating that the predicted probability output by the model has good calibration, that is, the predicted probability can accurately reflect the actual risk level.

[0035] Furthermore, this embodiment also employs decision curve analysis to evaluate the clinical net benefit of the combined predictive model. Decision curve analysis is a method for evaluating the clinical applicability of a predictive model. It determines whether the model can provide additional value for clinical decision-making by comparing the net benefit of model-guided decisions with the net benefit of the default strategy (i.e., intervention for all patients or no intervention) at different threshold probabilities. Figure 6 The results of decision curve analysis (DCA) of the joint prediction model are presented. The horizontal axis represents the high-risk threshold, indicating the minimum threshold at which clinicians decide to take intervention based on the model's predicted probability. The vertical axis represents the net benefit, quantifying the net clinical benefit of using the model to guide decision-making at the corresponding threshold compared to the baseline strategy. The None curve represents the baseline strategy of not intervening in any patients, with a net benefit of zero; the All curve represents the baseline strategy of intervening in all patients. The Train curve and Val curve represent the net benefit curves of the joint prediction model of this invention on the training and validation datasets, respectively. Figure 6 As can be seen, on the training dataset, within a wide range of threshold probabilities from 0.1 to 1.0, the model's net benefit curve lies above the curves for both the "full intervention" and "no intervention" extreme strategies, indicating that the model can provide positive clinical net benefit within this threshold range. On the validation dataset, within a range of threshold probabilities from 0.2 to 0.68, the model's net benefit curve also outperforms the extreme strategies. This result further supports the practicality and reliability of the method of this invention in real-world clinical applications.

[0036] In summary, this invention employs a three-tiered screening mechanism—comprising correlation elimination, LASSO sparsity compression, and Bootstrap resampling stability scoring—in the radiomics feature selection stage. This mechanism not only removes redundant features but also identifies a subset of target features with high cross-sample stability, effectively reducing overfitting risk and improving the model's generalization performance. In the information fusion stage, the introduction of a product interaction term between radiomics scores and clinical scores enables the fusion function to characterize the synergistic effect between the two information sources. Furthermore, the introduction of a nonlinear transformation term enhances the representation of nonlinear information in radiomics scores, allowing for a more comprehensive exploration of complementary relationships between multiple information sources compared to a single linear weighted fusion method. Finally, in the region of interest acquisition stage, a seed-point-based semi-automatic segmentation method is used, simplifying the physician's operation to two steps: single-point selection and boundary confirmation. This effectively shortens the processing time for a single data instance and reduces segmentation differences between different operators. The data from the examples show that the above three-level screening mechanism reduces the initial 1923 features to 8. The joint prediction model achieves an AUC of 0.92 on the validation dataset. Both calibration curve and decision curve analyses show good predictive consistency and clinical net benefit, verifying the comprehensive improvement effect of the technical solution of the present invention in terms of prediction stability, accuracy and processing efficiency.

[0037] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0038] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0039] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0040] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A radiomics and clinical nonlinear fusion prediction method, characterized in that, Including the following steps: S1: Obtain the acquired prostate dual-parameter magnetic resonance image data of the object to be processed, the data including T2-weighted imaging image data and diffusion-weighted imaging image data; S2: Based on the segmentation initialization information provided by the user, the T2-weighted imaging image data and diffusion-weighted imaging image data are processed using a semi-automatic segmentation method to obtain the region of interest; S3: By extracting radiomics features from the region of interest, a high-dimensional initial set of radiomics features is obtained; S4: Obtain a subset of target image omics features by performing hierarchical screening on the initial image omics feature set. The hierarchical screening includes feature correlation elimination, LASSO sparsity compression and Bootstrap resampling stability score screening performed sequentially. S5: Based on a subset of target image omics features, an image omics score is calculated using a pre-defined scoring model. S6: Obtain the existing clinical feature data of the object to be processed, and calculate the clinical score based on the clinical feature data through a preset clinical scoring model. S7: Based on a nonlinear weighted fusion function that includes a product interaction term of radiomics scores and clinical scores, the radiomics scores and clinical scores are fused to calculate a joint predictive score. S8: Output the joint prediction score and its visualization results as auxiliary prediction data related to the expression level of Ki-67 in the object to be processed.

2. The radiomics and clinical nonlinear fusion prediction method as described in claim 1, characterized in that, In step S2, the user-provided segmentation initialization information is the seed point coordinate data of the lesion center region, and the semi-automatic segmentation method is the region growing algorithm.

3. The radiomics and clinical nonlinear fusion prediction method as described in claim 2, characterized in that, The region growth algorithm is as follows: according to the preset neighborhood connectivity criterion and gray-level difference threshold, neighboring pixels whose gray-level difference with the seed point is less than the gray-level difference threshold are included in the growth region until the preset termination condition is met.

4. The radiomics and clinical nonlinear fusion prediction method as described in claim 1, characterized in that, In step S4, feature correlation elimination specifically involves: calculating the correlation coefficient between any two features in the initial radiomics feature set; when the absolute value of the correlation coefficient exceeds a preset threshold, eliminating features with weak correlation to the prediction target to obtain a subset of candidate radiomics features.

5. The radiomics and clinical nonlinear fusion prediction method as described in claim 4, characterized in that, In step S4, LASSO sparsity compression specifically involves: constructing a LASSO regression model for a subset of candidate image omics features, using cross-validation to determine the optimal penalty coefficient, and removing features whose regression coefficients are compressed to zero.

6. The radiomics and clinical nonlinear fusion prediction method as described in claim 5, characterized in that, In step S4, the Bootstrap resampling stability score selection is specifically as follows: the training data is resampled multiple times with replacement, and LASSO sparsity compression is repeatedly performed after each resampling. The frequency ratio of each candidate feature being selected is counted as the stability score, and only features with stability scores not lower than a preset threshold are retained as the target image omics feature subset.

7. The radiomics and clinical nonlinear fusion prediction method as described in claim 1, characterized in that, In step S6, the clinical characteristic data include at least one of the following: PI-RADS score, Gleason score, and prostate-specific antigen level.

8. The radiomics and clinical nonlinear fusion prediction method as described in claim 1, characterized in that, In step S7, the nonlinear weighted fusion function further includes a nonlinear transformation term based on radiomics scoring, and the fusion weight coefficients of the product interaction term and the nonlinear transformation term are determined by optimization on the training dataset using a combination of multivariate logistic regression and grid search.

9. The radiomics and clinical nonlinear fusion prediction method as described in claim 8, characterized in that, The nonlinear transformation term corresponds to the Sigmoid function, which is used to map the radiomics score to a preset numerical range, thereby enhancing the ability of the joint prediction score to represent the nonlinear information in the radiomics score.

Images