A tumor imaging prediction system based on IVIM quantitative parameter analysis

Abstract

The present invention relates to a tumor imaging genomics prediction system based on IVIM quantitative parameter analysis, comprising: an image acquisition module for acquiring different b The DWI images of the target part are formed by using the image registration module. b The DWI image with a value of 0 is used as a reference to align all DWI images; the image processing module inputs the aligned DWI image set into the IVIM parameter estimation model to obtain the IVIM quantitative parameter estimation value, and obtains the IVIM parameter image set according to the IVIM quantitative parameter estimation value; the lesion prediction module obtains the disease prediction result of the target part of the patient through the SVM prediction model based on the IVIM parameter image set. The above system solution can achieve fast and more accurate IVIM parameter estimation, truly reflect the movement of water molecules in the tumor, reflect tumor heterogeneity, and combine impactomics for accurate prediction.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a tumor imaging genomics prediction system based on IVIM quantitative parameter analysis. Background Art

[0002] Radiomics has been widely proven to be an effective tool for analyzing tumor heterogeneity and predicting disease-related classification problems. It can extract shape features, first-order features, and texture features from images of selected areas using a variety of operators. These features are screened using different statistical methods, and disease prediction is achieved by building classification models. It has been proven to be effective in predicting multiple diseases and has important clinical value. Diffusion-weighted imaging (DWI) is a non-invasive magnetic resonance functional imaging technology that does not rely on contrast agents. It can fit data through incoherent motion within voxels to generate parametric images to reflect tumor heterogeneity.

[0003] Studies have shown that it is feasible to use physical information networks (PINNs) for IVIM quantitative parameter estimation. However, because the currently used PINNs use multi-layer perceptrons to fit complex functional relationships, the estimated results are negatively affected by the catastrophic forgetting of the multi-layer perceptrons and the poor performance of the multi-layer perceptrons in fitting nonlinear functions, resulting in the inability to perform accurate voxel-level analysis. In addition, due to the limited estimation accuracy of the current IVIM fitting method, it is very difficult to perform voxel-level DWI analysis on parameter images. At the same time, the low estimation accuracy greatly affects the effectiveness of imaging omics feature extraction.

[0004] In addition, when estimating IVIM quantitative parameters, the fitting cannot be converted into a linear problem, and it is more difficult to obtain accurate quantitative parameters when performing voxel-level analysis because of the greater image noise. Previous fitting often uses the least squares method, Bayesian estimation method and neural network method. Among them, the neural network method shows great potential in estimation performance, but it is still difficult to make accurate estimates under the influence of noise. At the same time, judging the molecular subtype of the tumor by combining multiple heterogeneous parameter images requires extremely high diagnostic experience from the imaging physician, and the process is easily affected by subjective judgment. Due to the lack of corresponding guidelines, even though many domestic centers have already used DWI as a routine sequence for tumor scanning, clinical workers often cannot fully and effectively utilize the collected data. Summary of the invention

[0005] The purpose of the present invention is to improve the existing physical information network (PINN) in the prior art, which is susceptible to catastrophic forgetting of multi-layer perceptrons and poor fitting performance of multi-layer perceptrons for nonlinear functions when estimating IVIM quantitative parameters. Therefore, it is difficult to support accurate voxel-level analysis of tumor images, resulting in clinical workers often being unable to fully and effectively utilize the collected patient DWI data. Based on this, this solution provides a tumor imaging genomics prediction system based on IVIM quantitative parameter analysis.

[0006] In order to solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0007] As a first aspect of the present application, the present application provides a tumor imaging omics prediction system based on IVIM quantitative parameter analysis, comprising:

[0008] The image acquisition module is used to collect images of the patient's target parts at different b The DWI images under the values ​​are formed into a DWI image set of the target part; b The value is the diffusion sensitivity factor;

[0009] The image registration module is connected to the image acquisition module signal to b The DWI image with a value of 0 was used as a reference to register all DWI images;

[0010] An image processing module is connected to the image registration module by signal. The image processing module has an IVIM parameter estimation model built in. The registered DWI image set is input into the IVIM parameter estimation model to obtain IVIM quantitative parameter estimation values, and an IVIM parameter image set is obtained according to the IVIM quantitative parameter estimation values;

[0011] The lesion prediction module is connected to the image processing module signal. The lesion prediction module has a built-in SVM prediction model. The lesion prediction result of the target part of the patient is obtained through the SVM prediction model according to the IVIM parameter image set.

[0012] Furthermore, the IVIM parameter estimation model is implemented based on a physical information neural network improved by a spline function, and the IVIM parameter estimation model includes an input layer, a hidden layer, and an output layer; wherein, a plurality of hidden layers are provided, and each hidden layer is connected in sequence;

[0013] The input layer is used to receive different b The DWI image set after value registration is used, and nonlinear calculation is performed on the voxel values ​​in the DWI image set and the result is output to the hidden layer;

[0014] The hidden layer is used for each hidden layer to extract abstract features from the output results of neurons in the previous layer;

[0015] The output layer is used to constrain the output of the hidden layer using Sigmoid as an activation function and output IVIM quantitative parameter estimation values.

[0016] Furthermore, in the IVIM parameter estimation model, an improved spline function is introduced as an activation function on the “edge” connecting each layer of neurons. The expression of the activation function is:

[0017] ;

[0018] Among them, q is the update weight; f(X) is the nonlinear activation function silu; spline(X) is the linear combination of multiple B-splines; X is the output of the neuron in the previous layer.

[0019] Furthermore, the output layer introduces a Sigmoid function to calculate and obtain IVIM quantitative parameters, and the expression of the estimated value of the IVIM quantitative parameter is:

[0020] ,

[0021] ,

[0022] ,

[0023] Wherein, Dt', Dp' and Fp' are the output IVIM quantitative parameter estimates, which are respectively expressed as the pure diffusion coefficient estimate of Dt', the perfusion coefficient estimate of Dp' and the perfusion fraction estimate of Fp'; Sigmoid (L) is an activation function used to constrain the range of the IVIM quantitative parameter estimate, and L is the output of the output layer neuron; Dt max , Dt min are the maximum and minimum values ​​of the pure diffusion coefficient estimation range; Dp max 、Dp min are the maximum and minimum values ​​of the perfusion coefficient estimation range; Fp max , Fp min are the maximum and minimum values ​​of the perfusion fraction estimation range, respectively.

[0024] Furthermore, the lesion prediction module includes a lesion area acquisition unit, a feature extraction unit, a feature splicing unit, an invalid feature elimination unit and a symptom prediction unit:

[0025] The lesion region acquisition unit is used to acquire the lesion region of the IVIM parameter image set;

[0026] The feature extraction unit is used to extract multiple imaging omics features of the lesion area from the IVIM parameter image set;

[0027] The feature stitching unit is used to stitch different radiomics features of the IVIM parameter image set to obtain the total radiomics features based on the IVIM parameter image;

[0028] The invalid feature elimination unit is used to eliminate invalid features of the overall imaging omics features;

[0029] The disease prediction unit is used to input the total imaging genomics features with invalid features eliminated into the SVM prediction model, so as to obtain the disease prediction result of the target part of the patient.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. By reasonably modifying the spline function to the mspline function, the fitting performance of the model is improved and the catastrophic forgetting problem of the multi-layer perceptron is avoided;

[0032] 2. Through the IVIM parameter estimation model, rapid and more accurate IVIM parameter estimation can be achieved on new data, which can truly reflect the movement of water molecules in the tumor and reflect tumor heterogeneity;

[0033] 3. Accurate prediction of tumor subtypes can be achieved through the impactomics analysis module in the lesion prediction module. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 is an exemplary structural diagram of a tumor imaging omics prediction system based on IVIM quantitative parameter analysis described in the embodiments;

[0036] Figure 2 is a schematic flow chart of the DWI image processing process of the tumor imaging omics prediction system based on IVIM quantitative parameter analysis described in the embodiment;

[0037] Figure 3 is an exemplary model structure diagram of the IVIM parameter estimation model described in the embodiments;

[0038] Figure 4 Schematic diagram of IVIM parameter images obtained through experimental comparison described in the embodiments;

[0039] Figure 5 It is a schematic diagram of the verification of the results obtained based on the ROC curve described in the embodiment. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0041] Figure 1 FIG. 4 is an exemplary structural diagram of a tumor imaging omics prediction system based on IVIM quantitative parameter analysis described in the embodiments, as shown in FIG. Figure 1 As shown, the tumor imaging genomics prediction system based on IVIM quantitative parameter analysis includes: an image acquisition module 110 for acquiring different images of the patient's target site. b The DWI images of the target part are formed by using the DWI images of the target part. b The image registration module 120 is connected to the image acquisition module 110 signal to b All DWI images are registered with reference to the DWI image with a value of 0; an image processing module 130, the image processing module has a built-in IVIM parameter estimation model, the registered DWI image set is input into the IVIM parameter estimation model to obtain IVIM quantitative parameter estimation values, and an IVIM parameter image set is obtained according to the IVIM quantitative parameter estimation values; a lesion prediction module 140 is connected to the image processing module signal, the lesion prediction module has a built-in SVM prediction model, and the disease prediction result of the target part of the patient is obtained through the SVM prediction model according to the IVIM parameter image set.

[0042] In some embodiments, the image acquisition module 110 may use a magnetic resonance scanner to scan the target part of the patient and acquire different b The DWI images of the target part are formed by taking the DWI images of the target part. Figure 2 The (a) in the formula indicates different b DWI image set with the highest value. DWI image, or diffusion-weighted imaging, is the only method that can measure the diffusion of water molecules in living tissues, that is, a method used to detect the diffusion movement of water molecules in human tissues. bThe value is the diffusion sensitivity factor of diffusion-weighted imaging. b Under the setting of the value, DWI images with different signal intensities can be measured. b The value will affect the sensitivity and signal-to-noise ratio of DWI images. Clinicians can set multiple groups of values ​​based on the scanned area, clinical information of the patient, and other data. b value to obtain different b DWI images at the following values.

[0043] Specifically, taking the patient's target site as the breast as an example, in order to determine whether the patient has triple-negative breast cancer, clinicians use a magnetic resonance imaging scanner to scan the patient's target site and collect different b The DWI images with the same value are collected to form a DWI image set. In order to quickly obtain the best DWI image effect, the following scanning parameter configuration is used during the acquisition process: the image of the target part is acquired using a magnetic resonance imaging instrument with a field strength of 1.5T or 3T. Preferably, a field strength of 3T can be selected to obtain a higher signal-to-noise ratio; the imaging sequence of the DWI image is SS-EPI, and the number of coil turns is set to 16; b The number of values ​​is greater than 5; b The value collection intervals include 0-50 and 600-800, and at least one value must be collected in each collection interval. b Value; each b The signal average times of the value are recommended to be greater than 3; the acquisition image layer thickness is not greater than 5mm; the plane resolution of the acquisition image is not greater than 2.5mm. The above settings can ensure the imaging quality of DWI images, the acquisition area can cover the level where the tumor is located, and the acquisition can be completed quickly in a shorter imaging time.

[0044] In some embodiments, the image registration module 120 is connected to the image acquisition module 110 for signal matching. b The DWI images with the same value are registered. b The DWI image with a value of 0 is used as the reference. b The DWI images of the highest value are translated, rotated or scaled so that all DWI images can be matched.

[0045] Specifically, the image registration module 120 performs the b When registering the DWI images of the values, perform the following steps:

[0046] Assume that the DWI image I is b The reference image with a value of 0, the DWI image J is b For the image to be registered whose value is not 0, the DWI image J is translated, rotated or scaled so that the DWI image I can be matched with the DWI image J, and the pixel points of the DWI image J are calculated as shown in the following formula;

[0047] ;

[0048] ;

[0049] x and y are the horizontal and vertical coordinates of the pixel point in DWI image I, respectively; x' and y' are the horizontal and vertical coordinates of the pixel point in DWI image J corresponding to the pixel point (x, y) in DWI image I after translation, rotation or scaling; θ is the rotation angle of DWI image J, tx and ty are the translation parameters of the pixel point (x', y') in DWI image J on the horizontal and vertical coordinates, respectively; s represents the scaling parameter.

[0050] Furthermore, when performing the registration process described above for all DWI images with different b values, a minimization optimization process is required, as shown in the following formula:

[0051] ,

[0052] Where min is the symbol for minimization, s, θ and T are scaling parameters, rotation angles and translation parameters respectively, Q is the number of all pixels in the DWI image, I represents the pixel coordinates (x, y) of DWI image I, and J represents the pixel coordinates (x', y') of DWI image J.

[0053] like Figure 2 As shown, it shows a schematic flow chart of the DWI image processing process of the tumor imaging prediction system based on IVIM quantitative parameter analysis. Figure 2 In ( b ) represents the DWI image after registration, where the area outlined by lines is the tumor contour, and the green outline line 21 represents b The reference DWI image when the value is 0, the outlined area of ​​the blue outline line 22 represents the image to be registered, and the outlined area of ​​the pink outline line 23 represents the image after registration.

[0054] In some embodiments, the image processing module 130 is signal-connected to the image registration module 120. The image processing module has a built-in IVIM parameter estimation model. The registered DWI image set is input into the IVIM parameter estimation model to obtain IVIM quantitative parameter estimation values, and an IVIM parameter image set is obtained based on the IVIM quantitative parameter estimation values.

[0055] In the prior art, intravoxel incoherent motion is a diffusion-weighted imaging derivative model (IVIM), which is an imaging method used to describe the microscopic motion of voxels. The IVIM quantitative parameters are used to evaluate the diffusion motion component and the blood perfusion component respectively, and the motion of water molecules in the voxel is divided into simple diffusion and perfusion components. Specifically, the function expression describing the IVIM quantitative parameters is:

[0056] ,

[0057] Among them, Dp, Fp, and Dt are IVIM quantitative parameters. Specifically, Dp is the perfusion coefficient, which reflects the perfusion-related components; Fp is the perfusion fraction, which reflects the percentage of the perfusion component in the total diffusion; Dt is the pure diffusion coefficient, which reflects the simple diffusion component of water molecules in the tissue, and S b To set b The diffusion weighted signal of the DWI image (i.e., the value determined in the acquisition interval) b The diffusion weighted signal of the DWI image corresponding to the value), S 0 is the diffusion weighted signal of the unweighted DWI image (i.e. b The diffusion weighted signal of the DWI image has a value of 0); b is the diffusion sensitivity factor; e is a natural constant.

[0058] In the embodiment described in the present application, the IVIM quantitative parameter estimation value is determined by the IVIM parameter estimation model, and the IVIM parameter image set is further obtained according to the IVIM quantitative parameter estimation value. The IVIM parameter estimation model is implemented based on a physical information neural network improved by a spline function, and includes an input layer, a hidden layer and an output layer; the hidden layer is provided with a plurality of layers, and each hidden layer is connected in sequence.

[0059] Figure 3 : is an exemplary model structure diagram of the IVIM parameter estimation model described in the embodiment. Figure 3 As shown in FIG. 1 , each layer of the input layer, hidden layer and output layer in the IVIM parameter estimation model contains multiple neurons, wherein the number of neurons in the input layer is determined by b The number of values ​​is determined; the hidden layer may include two fully connected hidden layers, and the number of neurons in each hidden layer may be different. In the present embodiment, the first hidden layer may be set to 3 neurons, and the second hidden layer may be set to 5 neurons; the number of neurons in the output layer is determined according to the number of IVIM quantitative parameters. In the present embodiment, the number of neurons in the output layer is 3.

[0060] In order to avoid the loss of features due to compression during operation, the number of neurons in a single hidden layer cannot be less than the number of IVIM quantitative parameters. Using a modified spline activation function in the hidden layer will have better fitting performance when dealing with nonlinear problems, but because nonlinear calculations will increase the complexity of the inverse optimization process, the following strategies can be used when designing neurons in the hidden layer: (1) Reduce the number of neurons and hidden layers to ensure acceptable training time, but at least two or more hidden layers should be included to ensure network performance; (2) Considering that it is more difficult to extract deep features, the number of neurons in different hidden layers needs to be gradually increased.

[0061] like Figure 3 As shown, in the IVIM parameter estimation model, an improved spline function is introduced as the edge function mspline(X) on the "edge" connecting each layer of neurons. The edge between each neuron is designed to allow reverse update through the improved spline function. The expression of the edge function mspline(X) is:

[0062] ,

[0063] Among them, q is the first model parameter, which is used to update the IVIM parameter estimation model; f(X) is the nonlinear activation function silu, which is expressed as ; spline(X) is a linear combination of multiple B-splines, expressed as ; Where X is the output of the neuron in the previous layer; e is a constant; Ci is the second model parameter, which is also the combination coefficient, and Bi(x) is the spline function. , n is the order of the spline function.

[0064] In practical applications, the more complex the model, the better the fitting effect, but the calculation and training time will also increase. For example, in the calculation of DWI images for patients with triple-negative breast cancer, the spline order can be selected as 3.

[0065] In the prior art, each neuron in a neural network performs an optimizable linear transformation. In this embodiment, the neuron directly accumulates the output of the neurons in the previous layer after edge function calculation. The neuron can be expressed as:

[0066] ,

[0067] Among them, Neuron(Y) is a neuron, Yi is the output of the previous layer of neurons calculated by the edge function, , W is the number of neurons in the previous layer.

[0068] In some embodiments, the input layer is used to receive differentb The DWI image set after value registration is used, and nonlinear calculation is performed on the voxel values ​​in the DWI image set and the results are output to the hidden layer; each hidden layer extracts abstract features from the output results of the neurons in the previous layer, and finally the calculation results are output by the output layer; in the output layer, Sigmoid is used as an activation function to constrain the output of the hidden layer, and the IVIM quantitative parameter estimation value is output.

[0069] In some embodiments, the output layer: uses Sigmoid as an activation function to constrain the output of the output layer neurons, and outputs IVIM quantitative parameter estimates.

[0070] The expression of the IVIM quantitative parameter estimate is:

[0071] ,

[0072] ,

[0073] ,

[0074] Wherein, Dt', Dp' and Fp' are the output IVIM quantitative parameter estimates, which are respectively expressed as the pure diffusion coefficient estimate of Dt', the perfusion coefficient estimate of Dp' and the perfusion fraction estimate of Fp'; Sigmoid (L) is an activation function used to constrain the range of the IVIM quantitative parameter estimate, and L is the output of the output layer neuron; Dt max , Dt min are the maximum and minimum values ​​of the pure diffusion coefficient estimation range; Dp max 、Dp min are the maximum and minimum values ​​of the perfusion coefficient estimation range; Fp max , Fp min are the maximum and minimum values ​​of the perfusion fraction estimation range, respectively.

[0075] In some embodiments, the range of IVIM quantitative parameter estimation values ​​is determined based on manually input values ​​by clinicians, that is, the maximum and minimum values ​​of the pure diffusion coefficient estimation range, perfusion coefficient estimation range, and perfusion fraction estimation range can be obtained based on user input.

[0076] In some embodiments, according to the obtained IVIM quantitative parameter estimation values ​​Dt', Dp' and Fp', the physical information function is used to obtain the fitting data of the DWI image set, thereby generating the IVIM parameter image corresponding to the target part. mspline ( b ) is:

[0077] ,

[0078] Where Dt' is the estimated value of pure diffusion coefficient; Dp' is the estimated value of perfusion coefficient; Fp' is the estimated value of perfusion fraction); b Corresponding to the DWI image b value; e is a constant; PINN mspline ( b ) outputs the fitting data of the DWI image set.

[0079] Furthermore, the corresponding IVIM parameter image is synthesized based on the fitting data of the DWI image set. Figure 2 The image shown in (c) represents the IVIM parameter image.

[0080] In some embodiments, the IVIM parameter estimation model is obtained through initial IVIM parameter estimation model training, and the training method of the IVIM parameter estimation model includes the following steps: using the Kaiming initialization method to initialize the weight parameters in the IVIM parameter estimation model to obtain the initial IVIM parameter estimation model; using a pre-prepared DWI image training set to input the initial IVIM parameter estimation model to obtain the fitting data of the DWI image training set; constructing a loss function to optimize the training based on the fitting data of the DWI image training set, and when the loss function result meets the preset conditions, the training ends and the IVIM parameter estimation model is obtained.

[0081] Specifically, the loss function is expressed as:

[0082] ,

[0083] Among them, S b and S 0 Set b The diffusion weighted signal of the DWI image under the value and the diffusion weighted signal of the DWI image without weighting; M is b Number of values; PINN mspline ( b ) is the fitting data of the DWI image training set output by the initial IVIM parameter estimation model.

[0084] In some embodiments, the initial IVIM parameter estimation model training process is a process for narrowing the gap between the fitting data and the real data of the DWI image, wherein the fitting data is calculated based on the IVIM quantitative parameter estimation values ​​obtained by the IVIM parameter estimation model implemented in this embodiment.

[0085] Furthermore, in order to make the IVIM parameter estimation model converge faster, the ADAM optimizer is used to optimize the first model parameter and the second model parameter of the IVIM parameter estimation model, wherein the first model parameter is q in the variable function, and the second model parameter is the combination coefficient Ci in the variable function. , n is the order of the spline function.

[0086] In some embodiments, the lesion prediction module 140 is connected to the image processing module 130 by signal. The lesion prediction module has a built-in SVM prediction model, and the symptom prediction result of the target part of the patient is obtained through the SVM prediction model based on the IVIM parameter image. The lesion prediction module 140 includes a lesion area acquisition unit, a feature extraction unit, a feature splicing unit, an invalid feature elimination unit and a symptom prediction unit:

[0087] The lesion region acquisition unit is used to acquire the lesion region of the IVIM parameter image set. In practical applications, after obtaining the IVIM parameter image set, clinical workers can b The lesion area was outlined layer by layer on the DWI image.

[0088] The feature extraction unit is used to extract multiple radiomic features of the lesion area in the IVIM parameter image set; specifically, the radiomic features include shape features (Feature 1 ), first-order features 2 ) and texture features 3 The IVIM parameter image set includes the corresponding pure diffusion coefficient image, perfusion coefficient image and perfusion fraction image, and the corresponding radiomics features are extracted from each IVIM parameter image. For example, the radiomics features extracted from the pure diffusion coefficient (Dt') image include shape features (Feature Dt’-1 ), first-order features Dt’-2 ) and texture features Dt’-3 ); Extract the radiomics features of the perfusion coefficient (Dp') image including shape features (Feature Dp’-1 ), first-order features Dp’-2 ) and texture features Dp’-3 ); Extract radiomic features of perfusion fraction (Fp') images including shape features (Feature Fp’-1 ), first-order features Fp’-2 ) and texture features Fp’-3 In practical applications, the pyradiomics toolkit can be used to extract impactomics features.

[0089] The feature stitching unit is used to stitch different radiomics features of the IVIM parameter image set to obtain the total radiomics feature based on the IVIM parameter image set; further, the expression for stitching IVIM parameter images from different radiomics features is:

[0090] ,

[0091] Among them, Feature IVIM is the overall imaging genomics feature, Feature Dt’-n ,Feature Dp’-n ,Feature Fp’-n These are the radiomics features extracted from the pure diffusion coefficient (Dt') image, the perfusion coefficient (Dp') image, and the perfusion fraction (Fp') image. , Concat is the concatenation function.

[0092] The invalid feature elimination unit is used to eliminate invalid features of the total radiomics features. In some embodiments, invalid features refer to features that are irrelevant to clinical indicators and redundant. In practical applications, recursive feature elimination (RFE) can be used to select valid features. The RFE method has been proven to be a method that can maximize the use of data and existing model architectures and can effectively screen features.

[0093] The disease prediction unit is used to input the total imaging genomics features with invalid features eliminated into the SVM prediction model, so as to obtain the disease prediction result of the target part of the patient.

[0094] In some embodiments, the SVM prediction model is implemented based on a support vector machine (SVM), which is an effective machine learning model that can achieve accurate classification by fitting the optimal hyperplane through a complex kernel function. Specifically, taking triple-negative breast cancer tumor prediction as an example, the SVM prediction model is trained in a data-driven manner by marking the presence / absence of triple-negative breast cancer with 0 and 1, respectively. The hyperplane finally formed in the feature space is used to distinguish whether the newly input feature belongs to the patient's breast cancer is triple-negative.

[0095] In some embodiments, the disease prediction results provided by the present application are compared with the parameter estimates obtained by the least squares method. Figure 4 Schematic diagram of IVIM parameter images obtained through experimental comparison described in the embodiment, as shown in FIG. Figure 4 As shown, Figure 4 (a) is an IVIM parameter image of the tumor site obtained by the method described in this application, Figure 4 In ( b) is a parameter image obtained by the least squares method. Furthermore, the IVIM parameter estimation model provided by the present invention is evaluated according to the ROC curve and the AUC value. The ‌ROC curve (Receiver Operating Characteristic Curve)‌, also known as the sensitivity curve, is a graphical tool for evaluating the performance of a binary classification model. The ROC curve shows the classification performance of the model under different thresholds by plotting the relationship between the true positive rate (TPR) and the false positive rate (FPR). The AUC value is the area under the ROC curve and is used to evaluate the performance of a binary classification model. The AUC value ranges between 0.5 and 1. The closer the AUC value is to 1, the higher the prediction accuracy of the model; conversely, the closer the AUC value is to 0.5, the lower the prediction accuracy of the model. Figure 5 As shown, Figure 5 ] are AUC values ​​of different parameter estimates obtained by the least squares method and the method provided in this application, wherein the light blue line 52 represents the AUC value (0.68) obtained by the least squares method, and the dark blue line 51 represents the AUC value (0.81) obtained by the system provided in this solution.

[0096] It can be proved from the above embodiments and the verification results of the system that the technical solution provided by the present application improves the fitting performance of the IVIM parameter estimation model by reasonably modifying the spline function to the mspline function, avoids the catastrophic forgetting problem of the multi-layer perceptron, and through the trained IVIM parameter estimation model, it is possible to quickly and more accurately obtain IVIM quantitative parameter estimates on new data, which can truly reflect the movement of water molecules in the tumor, reflect tumor heterogeneity, and thus achieve a more accurate prediction of the tumor condition.

[0097] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0098] If the concepts of modules, models, layers, etc. described in this solution are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0099] Those of ordinary skill in the art will appreciate that the modules of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the constituent modules and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0101] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A tumor imaging prediction system based on IVIM quantitative parameter analysis, characterized in that: include: The image acquisition module is used to collect images of the patient's target parts at different b The DWI images under the values ​​are used to form a DWI image set of the target part; Said b The value is the diffusion sensitivity factor; The image registration module is connected to the image acquisition module signal to b The DWI image with a value of 0 was used as a reference to register all DWI images; An image processing module is connected to the image registration module by signal. The image processing module has an IVIM parameter estimation model built in. The registered DWI image set is input into the IVIM parameter estimation model to obtain IVIM quantitative parameter estimation values, and an IVIM parameter image set is obtained according to the IVIM quantitative parameter estimation values; A lesion prediction module is connected to the image processing module by signal, and the lesion prediction module has a built-in SVM prediction model, and obtains the disease prediction result of the target part of the patient through the SVM prediction model according to the IVIM parameter image set; The IVIM parameter estimation model in the image processing module is implemented based on a physical information neural network improved by a spline function, and the IVIM parameter estimation model includes an input layer, a hidden layer and an output layer; wherein a plurality of hidden layers are provided, and each hidden layer is connected in sequence; The input layer is used to receive different b The DWI image set after value registration is used, and nonlinear calculation is performed on the voxel values ​​in the DWI image set and the result is output to the hidden layer; The hidden layer is used for each hidden layer to extract abstract features from the output results of neurons in the previous layer; The output layer is used to constrain the output of the hidden layer by using Sigmoid as an activation function, and output IVIM quantitative parameter estimation values; In the IVIM parameter estimation model, an improved spline function is introduced as an edge function on the "edge" connecting each layer of neurons. The expression of the edge function is: ， Wherein, q is the first model parameter, which is used to update the IVIM parameter estimation model; f(X) is the nonlinear activation function silu, which is expressed as ; spline(X) is a linear combination of multiple B-splines, expressed as ; Where X is the output of the neuron in the previous layer; e is a constant; Ci is the second model parameter, which is also the combination coefficient, and Bi(x) is the spline function. , n is the order of the spline function; The neurons in the IVIM parameter estimation model directly accumulate the outputs of the neurons in the previous layer after edge function calculation. The neurons can be expressed as: ， Among them, Neuron(Y) is a neuron, Yi is the output of the previous layer of neurons calculated by the edge function, , W is the number of neurons in the previous layer; The output layer in the IVIM parameter estimation model uses Sigmoid as an activation function to constrain the output of the output layer neurons and outputs IVIM quantitative parameter estimation values, which are expressed as: ， ， ， Wherein, Dt', Dp' and Fp' are the output IVIM quantitative parameter estimates, which are respectively expressed as the pure diffusion coefficient estimate of Dt', the perfusion coefficient estimate of Dp' and the perfusion fraction estimate of Fp'; Sigmoid (L) is an activation function used to constrain the range of the IVIM quantitative parameter estimate, and L is the output of the output layer neuron; Dt max , Dt min are the maximum and minimum values ​​of the pure diffusion coefficient estimation range; Dp max 、Dp min are the maximum and minimum values ​​of the perfusion coefficient estimation range; Fp max 、Fp min are the maximum and minimum values ​​of the perfusion fraction estimation range, respectively; The image processing module obtains the fitting data of the DWI image set using the physical information function according to the obtained IVIM quantitative parameter estimation value, thereby generating the IVIM parameter image set corresponding to the target part; wherein the physical information function PINN mspline ( b ) is: , Where Dt' is the estimated value of pure diffusion coefficient; Dp' is the estimated value of perfusion coefficient; Fp' is the estimated value of perfusion fraction; b Corresponding to the DWI image b value; e is a constant; PINN mspline ( b ) outputs the fitting data of the DWI image set.

2. The tumor imaging prediction system based on IVIM quantitative parameter analysis according to claim 1, characterized in that: The IVIM parameter estimation model training method comprises: Kaiming initialization method is used to initialize the weight parameters in the IVIM parameter estimation model to obtain the initial IVIM parameter estimation model; Using the pre-prepared DWI image training set to input the initial IVIM parameter estimation model to obtain fitting data of the DWI image training set; A loss function is constructed to optimize the training of the fitting data according to the DWI image training set, and when the loss function result meets the preset conditions, the training is completed to obtain the IVIM parameter estimation model.

3. The tumor imaging prediction system based on IVIM quantitative parameter analysis according to claim 1, characterized in that: The lesion prediction module includes a lesion area acquisition unit, a feature extraction unit, a feature splicing unit, an invalid feature elimination unit and a symptom prediction unit: The lesion region acquisition unit is used to acquire the lesion region of the IVIM parameter image set; The feature extraction unit is used to extract multiple imaging omics features of the lesion area from the IVIM parameter image set; The feature stitching unit is used to stitch different radiomics features of the IVIM parameter image set to obtain the total radiomics features based on the IVIM parameter image; The invalid feature elimination unit is used to eliminate invalid features of the overall imaging omics features; The disease prediction unit is used to input the total imaging genomics features with invalid features eliminated into the SVM prediction model, so as to obtain the disease prediction result of the target part of the patient.

4. The tumor imaging prediction system based on IVIM quantitative parameter analysis according to claim 3, characterized in that: The imaging omics features include shape features, first-order features and texture features.

Images