Method, system and product for predicting unilateral renal function grading based on CT subtraction imaging

Abstract

The invention discloses a method, a system and a product for predicting unilateral renal function grading based on CT subtraction imaging, and the method comprises the steps: firstly, extracting a cortical medullary enhancement period CTU image of a patient, and obtaining a plain-scan CTU image and an enhanced CTU image; then aligning the plain-scan CTU image and the enhanced CTU image; automatically segmenting the bilateral renal cortex by using a neural network model, and further reasoning to obtain a renal cortex mask; constructing a subtraction imaging mode by pixel-by-pixel subtraction based on the registered enhanced image and the plain-scanned image; then extracting subtraction imaging modal radiation group features to obtain unilateral kidney radiomics features; and finally, predicting the kidney function based on the kidney function prediction model. According to the method, the recognition precision of the kidney cortex area is improved, interference features are eliminated, the signal-to-noise ratio is improved, and the kidney function of the patient with total renal insufficiency is accurately and noninvasively predicted. Valuable guidance can be provided for clinical decision making, and clinicians can be helped to better evaluate patients suffering from total renal insufficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and relates to a unilateral kidney function grading method, system and product, in particular to a CT kidney key structure semantic segmentation, subtraction imaging, imageomics feature extraction and data modeling method, system and product based on artificial intelligence. BACKGROUND

[0002] Acute and chronic renal dysfunction has become a public health problem threatening the world. Early detection of renal impairment and treatment is the key to delaying the development of chronic renal failure. Glomerular filtration rate (GFR) is the best indicator of renal function assessment. In clinical practice, the assessment of split kidney GFR is needed for the decline in renal function caused by various reasons, which is of great significance for treatment decision making and even surgical guidance. Imaging diagnosis plays a core role in the assessment of various renal dysfunction. Among various imaging techniques, multi-phase enhanced computed tomography urography (CTU) is widely used because it can accurately locate the position and severity of urinary obstruction through imaging stages such as plain scan, cortical medullary phase, renal parenchymal phase and excretion phase.

[0003] Emission computed tomography (ECT), also known as renal dynamic imaging, is the gold standard for assessing renal function, but its clinical application is limited due to its invasiveness and dependence on radioactive drugs, especially in patients with impaired renal function. In addition to the additional economic burden, ECT requires intravenous injection of radioactive drugs for imaging, and conditions such as contrast agent allergy and increased renal function burden prevent many patients from being able to or willing to accept it.

[0004] The renal cortex is the key site for filtration and is extremely sensitive to structural and functional changes caused by various reasons. Conditions such as hydronephrosis or compression of adjacent lesions can cause the cortex to thin, and this thinning is closely related to the decline in renal function. Radiologists usually rely on the cortical morphology shown in CTU images to infer functional impairment. However, this assessment is highly dependent on individual experience, is highly subjective, and has poor inter-observer consistency. In recent years, deep learning techniques, particularly three-dimensional convolutional neural networks (3D CNN), have shown significant potential in medical imaging tasks. These technologies have shown good application prospects in organ segmentation, disease detection and functional analysis. However, many existing models are limited by small data sets, limited attention to clinical practicality, and failure to fully integrate into the actual diagnostic process. Considering the above problems, there is an urgent need for a non-invasive and reliable alternative method to assess renal function. SUMMARY

[0005] In order to solve the problem of evaluating the accuracy of split kidney function in patients with all-cause renal dysfunction in the prior art, the present application provides a CT subtraction imaging-based prediction method, system and product for unilateral kidney function grading based on CTU artificial intelligence kidney cortex precise segmentation and new modalities of subtraction imaging.

[0006] The technical scheme adopted by the method of the present application is: a CT subtraction imaging-based prediction method for unilateral kidney function grading, comprising the following steps: Step 1: extracting the patient's cortical medulla enhanced CTU image to obtain the plain CTU image and the enhanced CTU image; Step 2: based on physical space resampling and affine rigid registration, aligning the plain CTU image and the enhanced CTU image; Step 3: using a neural network model to automatically segment the bilateral kidney cortex, and further inferring to obtain the split kidney cortex mask; Step 4: based on the registered enhanced image and the plain image output in step 2, constructing a subtraction imaging modality by pixel-by-pixel subtraction; Step 5: extracting the radiological group features of the subtraction imaging modality to obtain the unilateral kidney image feature; Step 6: predicting the split kidney function based on a split kidney function prediction model.

[0007] As a preferred embodiment, the specific implementation of step 2 comprises the following sub-steps: Step 2.1: taking the enhanced CTU image as the reference sequence, creating a resampler according to the size, resolution, origin position and direction matrix of the enhanced CTU, and resampling the CT plain image sequence to the same coordinate space as the enhanced CT; Step 2.2: using linear interpolation resampling to process the CT grayscale image; setting the transformation to a unit transformation, only resampling the image to a new spatial framework; for pixels without corresponding original image positions in the new space, the default pixel value is set to 0; Step 2.3: unifying the image direction to left-posterior-superior, and initializing an affine transformation with the geometric center as the reference; Step 2.4: setting the registration strategy and multi-resolution strategy, performing registration to obtain the final transformation matrix; then mapping the plain CT image to the spatial framework of the enhanced CT according to the transformation.

[0008] As preferred, in step 2.4, the registration strategy uses MattesMutualInformation as the similarity measurement function, the optimizer is set to GradientDescent, and the learning rate, maximum number of iterations, convergence threshold and convergence window are set; the metric_record callback function is called to record the current metric value at each iteration to facilitate the analysis of the convergence process. The multi-resolution strategy uses a three-level pyramid registration to improve the stability and efficiency of the registration through downsampling and Gaussian smoothing processing.

[0009] As preferred, in step 3, first, the bilateral renal cortex structure is labeled layer by layer for the input image data, and the corresponding mask file is generated; then the non-zero region data is cropped; then the data is resampled and normalized, the resampling is used to normalize the interval of all images to the median interval of the data set, the original data is resampled using cubic spline interpolation, and the mask data is resampled using nearest neighbor interpolation method; finally, the neural network model is input to perform image segmentation, and further inference is performed to obtain the renal cortex mask; The neural network model is composed of 7 different stages, each stage is responsible for extracting different levels of features, gradually from low-level features to high-level features; the number of feature channels of each stage is [32, 64, 128, 256, 320, 320, 320] respectively; each stage uses 3D convolution to capture spatial information, the convolution kernel size of each stage is: Stage 1: [1, 3, 3]; Stage 2: [1, 3, 3]; Stage 3-7: [3, 3, 3]; the stride is set to: Stage 1: [1,1, 1]; Stage 2: [1, 2, 2]; Stage 3: [1, 2, 2]; Stage 4: [2, 2, 2]; Stage 5: [2, 2, 2]; Stage 6: [1, 2, 2]; Stage 7: [1, 2, 2]; the encoder and decoder have 2 convolution layers in each stage [2, 2, 2, 2, 2, 2, 2]; no Dropout layer is used; the nonlinear activation function specified for use is torch.nn.LeakyReLU, and the nonlinear activation function is calculated in place.

[0010] As preferred, in step 3, the probability map generated by the neural network model is converted to a binary mask through thresholding, and dilation and erosion operations are used to remove artifacts and noise smaller than the set threshold; the connected components in the segmentation result are analyzed to remove connected regions smaller than the set threshold; finally, the segmentation result is smoothed to output an accurate probability map.

[0011] As preferred, in step 3, the neural network model is a trained model; In the training process, the loss function used Combining Dice loss And cross-entropy loss ;

[0012] Where, : Dice loss function; K: class set; |K|: number of classes; k∈K: the kth class being calculated; I: position set of all pixels (or voxels) in the image; i∈I: the pixel (or voxel) position being calculated; u is the Softmax probability output, and v is the hard-coded ground truth; : The prediction value of the model for the ith pixel in the kth class; : The value of the ith pixel in the kth class in the real label. : The intersection part of the prediction and the real label in the kth class; : The union of the total prediction and actual number of the prediction and the real label in the kth class. : Cross-entropy loss value; Y: real label. P: prediction probability output by the model.

[0013] Learning rate adjustment strategy in the training process: use Adam optimizer, updated weight parameter ; Represents the convolution weight parameter before updating; The current learning rate; Represents the first moment estimate after bias correction; The second moment estimate after bias correction; ε is a small constant to prevent the denominator from being zero, maintaining numerical stability.

[0014] Stop training when the upper limit of the training period epoch is reached, or when the exponential moving average loss of the validation set decreases by no more than 5e-3 in 60 training periods epoch, or the learning rate is reduced to 1e-6; In the training process, a variety of data augmentation techniques are used, including random rotation, random scaling, random elastic transformation, gamma correction and mirroring, to improve the generalization ability of the model.

[0015] In the training process, in order to increase the stability of the network, when Patch sampling, more than 1 / 3 of the pixels in a batch are foreground class pixels.

[0016] As preferred, in step 4, firstly, the enhanced CT image after registration processing is subjected to a voxel-by-voxel subtraction operation with the corresponding CT image in the plain scan period; then, in order to remove minor differences and interference of artifacts, a tolerance threshold setting is introduced, and for all subtraction voxels, if the difference in the enhanced value is less than the threshold, it is considered to have no significant enhancement and is regarded as background.

[0017] As preferred, in step 6, the split kidney function prediction model adopts an ElasticNet regression model, and the reference objective function is as follows:

[0018] Wherein, y i is the actual observation value of the i th sample; x ij is the j th input feature value of the i th sample; n is the total number of samples; p is the total number of features; β 0 is the intercept of the model; β j is the weight coefficient of the j th feature, i.e., the model parameter; in the formula, is the regularization intensity coefficient, is the weight ratio of the L1 norm and the L2 norm regularization term; The split kidney function prediction model is a trained model; During the training process, the matching renal parenchyma dynamic phenomenon 99mTc-DTPA examination results of the patient are synchronously collected, and the GFR is calculated based on the 99mTc-DTPA renal dynamic imaging by using the Gates analysis method:

[0019]

[0020]

[0021] During the training process, a 5-fold cross-validation strategy is adopted for training, so as to ensure the robustness and generalization of the model.

[0022] The technical scheme of the system of the application is as follows: a split kidney function grading system based on CT subtraction imaging, comprising the following modules: An image acquisition and preprocessing module is used to extract the patient's cortical medulla enhanced CTU image, obtain the plain scan CTU image and the enhanced CTU image; A multi-modal CTU image registration module based on physical space resampling and affine rigid registration is used to align the plain scan CTU image and the enhanced CTU image based on physical space resampling and affine rigid registration; A segmentation module based on a neural network is used to automatically segment the bilateral renal cortex by using a neural network model, and further reasoning is performed to obtain a split renal cortex mask; A subtraction processing module is configured to construct a subtraction imaging modality by pixel-by-pixel subtraction based on the registered enhanced image and the plain scan image; A feature extraction module is configured to extract a radiation group feature of the subtraction imaging modality to obtain a single kidney image group feature. A split kidney function prediction model is constructed, and the split kidney function is predicted based on the split kidney function prediction model.

[0023] The technical scheme of the system of the present application is: a product for predicting single kidney function grading based on CT subtraction imaging, comprising computer program instructions, when the computer program instructions are run on a computer, the computer executes a method for predicting single kidney function grading based on CT subtraction imaging.

[0024] The present application combines multi-phase enhanced CTU imaging, affine rigid registration and advanced deep learning technology to construct a subtraction imaging modality, realizes automatic segmentation of high-precision renal cortex, and further extracts features based on image group, trains an ElasticNet regression model to predict split kidney function. This method not only improves the recognition accuracy of the renal cortex region, eliminates interference features, and improves the signal-to-noise ratio, but also accurately and non-invasively predicts the renal function of patients with all-cause renal dysfunction. It can provide valuable guidance for clinical decision-making and help clinicians better evaluate and all-cause renal dysfunction patients. BRIEF DESCRIPTION OF DRAWINGS

[0025] The technical scheme of the present application is further illustrated by using the following examples and specific embodiments. In addition, some drawings are also used in the process of explaining the technical scheme. For those skilled in the art, other drawings and the intention of the present application can be obtained without creative labor based on these drawings.

[0026] Figure 1 The method flowchart of the embodiment of the present application is shown in the figure; Figure 2 The multi-modal CTU data registration principle diagram of the embodiment of the present application is shown in the figure; Figure 3 The renal cortex segmentation model structure diagram of the embodiment of the present application is shown in the figure; Figure 4 The subtraction imaging modality generation result schematic diagram of the embodiment of the present application is shown in the figure, A is a subtraction calculation result schematic diagram, and B is a tolerance filtering result schematic diagram; Figure 5 The system structure block diagram of the embodiment of the present application is shown in the figure; Figure 6 The data verification fitting result of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0027] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0028] The embodiment of the present invention provides a method for predicting unilateral renal function classification based on CT subtraction imaging, the flow chart of which is as follows: Figure 1 As shown, the following steps are included: Step 1: Extract the patient's cortical and medullary enhancement phase CTU images to obtain plain scan CTU images and enhanced CTU images; In one embodiment, cortical and medullary enhancement-phase CTU images are acquired with a slice thickness of 5 mm and a resolution of 512 × 512 pixels and stored in DICOM format. Both plain and contrast-enhanced CT images are loaded using the float32 data type. During this process, the image data undergoes manual cleaning to ensure quality and reliability.

[0029] Step 2: Align the plain scan CTU image and the enhanced CTU image based on physical space resampling and affine rigid registration; In one embodiment, see Figure 2 , the specific implementation of step 2 includes the following sub-steps: Step 2.1: Spatial resampling based on the reference sequence; Step 2.1.1: Create a resampler; Using the enhanced CTU image as the reference sequence, a SimpleITK.ResampleImageFilter resampler is created based on the size, resolution, origin position, and orientation matrix of the enhanced CTU to resample the CT plain scan imaging sequence to the same coordinate space as the enhanced CT.

[0030] Step 2.1.2: Specify the interpolation method; Use linear interpolation (sitkLinear) to resample the CT grayscale image, that is, estimate the pixel intensity corresponding to the new spatial position through linear interpolation based on the pixel value of the original image.

[0031] Step 2.1.3: Transform settings; Set the transformation to a unit transform (sitk.Transform()). This does not transform the image content itself, such as rotation or translation, but only resamples the image to the new spatial frame. For pixels that do not correspond to the original image position in the new space (such as edge positions), a default pixel value of 0 is assigned.

[0032] Step 2.2: Align multimodal imaging data based on affine rigid registration; Step 2.2.1: Initial registration transformation setting; Load image Float32 type image data; use DICOMOrient to unify the image direction to LPS (left-posterior-superior), avoid registration error caused by inconsistent direction. Use CenteredTransformInitializer to initialize an affine transformation (AffineTransform) based on the geometric center, provide reasonable initial parameters for subsequent optimization.

[0033] Step 2.2.2: Set registration strategy; Use MattesMutualInformation as the similarity measure function, suitable for different image modalities or large gray contrast differences. Set the optimizer to GradientDescent, set the learning rate, maximum iteration number, convergence threshold and convergence window. Call the metric_record callback function to record the current metric value at each iteration, which facilitates the analysis of the convergence process.

[0034] Step 2.2.3: Multi-resolution strategy; Implement a three-level pyramid registration, improve the stability and efficiency of registration through downsampling and Gaussian smoothing processing.

[0035] Step 2.2.4: Execute registration and resampling; Use registration_method.Execute() to execute the registration process and get the final transformation matrix. Then call Resample to map the plain CT image to the enhanced CT spatial framework according to this transformation. Unify the multi-modal CTU image space.

[0036] Step 3: Use the neural network model to automatically segment the bilateral renal cortex, and further infer to obtain the renal cortex mask; In one embodiment, the specific implementation of step 3 includes the following sub-steps: Step 3.1: Based on the above image data, use 3D Slicer software to manually annotate the bilateral renal cortex structure layer by layer, respectively, to generate the corresponding mask file.

[0037] Step 3.2: Overall data cropping. In the data preprocessing stage, only the non-zero area is cropped, thereby reducing the consumption of computing resources.

[0038] Step 3.3: Resampling. For different voxel intervals existing in the data set, by default, all image intervals are normalized to the median interval of the data set. The original data is resampled using cubic spline interpolation, while the mask data uses nearest neighbor interpolation method.

[0039] Step 3.4: CT data normalization. The CT data of each patient is z-score normalized to standardize the data differences between different patients and enhance the robustness of the model.

[0040] Step 3.5: Constructing a neural network model to automatically segment the bilateral renal cortex; The neural network model structure is as follows Figure 3 , which consists of 7 different stages, each responsible for extracting different levels of features, gradually from low-level features to high-level features; the number of feature channels in each stage is [32, 64, 128, 256, 320, 320, 320] respectively; each stage uses 3D convolution (torch.nn.modules.conv.Conv3d) to effectively capture spatial information, and the kernel size of each stage is: Stage 1: [1, 3, 3]; Stage 2: [1, 3, 3]; Stage 3-7: [3, 3, 3]; the stride is set to: Stage 1: [1, 1, 1]; Stage 2: [1, 2, 2]; Stage 3: [1, 2, 2]; Stage 4: [2, 2, 2]; Stage 5: [2, 2, 2]; Stage 6: [1, 2, 2]; Stage 7: [1, 2, 2]; Each stage of the encoder and decoder has 2 convolutional layers [2, 2, 2, 2, 2, 2, 2]; No Dropout layer is used; the nonlinear activation function specified for use is torch.nn.LeakyReLU, which can effectively avoid the problem of gradient disappearance and enhance the nonlinear expression ability of the model. The nonlinear activation function is calculated in place to save memory. Deep supervision is set.

[0041] Step 3.6: Post-processing. Mainly using connected component analysis to further optimize the segmentation results and ensure the accuracy and consistency of the output. The generated probability map is converted to a binary mask (i.e. segmentation of background and foreground) through thresholding, and dilation and erosion operations are used to remove small artifacts and noise. The dilation operation can connect separate regions, while the erosion operation can remove small objects. Analyze the connected components in the segmentation results and remove connected regions smaller than a certain threshold (remove regions smaller than a certain volume). Smooth the segmentation results.

[0042] Step 3.7: Inference based on the trained network to generate a mask file of the bilateral renal cortex related structure, and obtain a unilateral kidney image from the median boundary segmentation mask.

[0043] In one embodiment, the neural network model is a trained model; Loss function used in training process Combined with Dice loss and cross-entropy loss ;

[0044] where, : Dice loss function; K: set of classes; |K|: number of classes; k∈K: the k-th class being computed; I: set of all pixel (or voxel) positions in the image; i∈I: the pixel (or voxel) position being computed; u is the Softmax probability output, and v is the hard-coded ground truth; : the model's prediction of the i-th pixel in the k-th class; : the value of the i-th pixel in the k-th class in the true label. : the intersection part of the prediction and the true label in the k-th class; : the union of the total predicted and actual number of the prediction and the true label in the k-th class. : cross-entropy loss value; Y: true label. P: predicted probability output by the model.

[0045] Learning rate adjustment strategy during training process: use Adam optimizer, initial learning rate set to 0.01, weight decay (L2 regularization) set to 0.00003 (3e-5); 250 iterations per training cycle epoch, 50 iterations per validation cycle, a total of 1000 cycles; updated weight parameters ; denotes the convolution weight parameters before updating; is the current learning rate; denotes the first moment estimate after bias correction; is the second moment estimate after bias correction; ε is a small constant (10-8) to prevent the denominator from being zero, maintaining numerical stability.

[0046] Stop training when the upper limit of the training cycle epoch is reached, or when the exponential moving average loss of the validation set decreases by no more than 5e-3 in 60 training cycles epoch, or the learning rate is reduced to 1e-6; During the training process, a variety of data augmentation techniques are used, including random rotation, random scaling, random elastic transformation, gamma correction, and mirroring, to improve the model's generalization ability.

[0047] During the training process, in order to increase the stability of the network, when Patch sampling, more than 1 / 3 of the pixels in a batch are foreground class pixels.

[0048] Step 4: Constructing a subtraction imaging modality based on the registered enhanced image and the plain scan image output in Step 2, pixel by pixel subtraction; To improve the model's recognition ability of enhanced structures in CTU images, subtraction imaging processing of the plain scan phase and the enhanced phase is adopted. In an embodiment, see Figure 4 The specific implementation of Step 4 includes the following sub-steps: A. Subtraction calculation; Perform voxel-wise subtraction operation on the registered enhanced phase CT image and the corresponding plain scan phase CT image, as follows: ; Through this subtraction operation, the signals of blood vessels, urinary system filling structures, and other enhanced related tissues are strengthened on the image, and the signals of background tissues (non-enhanced regions) are suppressed, further improving the signal-to-noise ratio of the lesion area and the surrounding normal tissues.

[0049] B. Tolerance filtering; To remove minor differences and artifacts, the system introduces a tolerance threshold setting: for all subtraction voxels, if the difference in their enhanced values is less than 80 HU, it is considered to have no significant enhancement and is treated as background.

[0050]

[0051] Such voxels are uniformly set to -1000 HU, simulating air values, to reduce interference.

[0052] Step 5: Extracting radiation group features from the subtraction imaging modality to obtain unilateral kidney image radiomics features; In an embodiment, after segmenting the renal cortex region, use the PyRadiomics package to extract radiation group features from the subtraction imaging mask. Extract 1280 features for each kidney image.

[0053] Step 6: Predicting split kidney function based on the split kidney function prediction model.

[0054] In an embodiment, the split kidney function prediction model uses an ElasticNet regression model with a regularization strength coefficient of 1.0 and a weight ratio of L1 norm and L2 norm regularization terms of 0.5. The reference objective function is

[0055] where y i is the actual observed value of the i-th sample; x ij is the j-th input feature value of the i-th sample; n is the total number of samples; p is the total number of features; β0 is the intercept (bias) of the model; β jis the weight coefficient of the jth feature, i.e., the model parameter; in the formula is the regularization intensity coefficient, is the weight ratio of the L1 norm and the L2 norm regularization term; The split kidney function prediction model is a trained model. During the training process, the patient's matched renal parenchyma dynamic phenomenon 99mTc-DTPA (99m technetium-diethylenetriaminepentaacetic acid) examination result is synchronously collected, and the GFR is calculated based on the 99mTc-DTPA renal dynamic imaging by using the Gates analysis method:

[0056]

[0057]

[0058] During the training process, 1280 features are used as input features, and the renal function in the renal dynamic imaging diagnosis report is used as the target output feature. The data set is divided into a training set and a test set in a ratio of 8:2, a 5-fold cross-validation strategy is used for training, and an ElasticNet regression model is trained. To ensure the robustness and generalization of the model.

[0059] See Figure 5 The embodiment also provides a prediction of unilateral kidney function grading system based on CT subtraction imaging, comprising the following modules: An image acquisition and preprocessing module is used to extract a patient's cortical medullary enhancement period CTU image, obtain a plain scan CTU image, and obtain an enhanced CTU image; A multi-modal CTU image registration module based on physical space resampling and affine rigid registration is used to align the plain scan CTU image and the enhanced CTU image based on physical space resampling and affine rigid registration; A segmentation module based on a neural network is used to automatically segment the bilateral kidney cortex by using a neural network model, and further infer to obtain a split kidney cortex mask; A subtraction processing module is used to construct a subtraction imaging modality by pixel-by-pixel subtraction based on the registered enhanced image and the plain scan image; A feature extraction module is used to extract a radio group feature of the subtraction imaging modality, and obtain a unilateral kidney image feature; A split kidney function prediction model is constructed, which is used to predict the split kidney function based on the split kidney function prediction model.

[0060] The embodiment also provides a prediction of unilateral kidney function grading product based on CT subtraction imaging, comprising computer program instructions, which, when running on a computer, enable the computer to execute the prediction of unilateral kidney function grading method based on CT subtraction imaging.

[0061] A total of 771 cases were included in the early stage of this project for data validation of kidney function prediction robustness. Based on the above process, the dual kidney features were extracted, and the split kidney function was predicted based on multi-sequence subtraction imaging. The regression model results showed that the split kidney GFR prediction value and the ECT gold standard regression correlation coefficient R 2 was 0.918 and 0.888, respectively, and the regression results are shown in Figure 6 . The AUC of the split kidney function classification based on the predicted value was 0.946 (95% confidence interval: 0.923-0.969), and the accuracy, sensitivity and specificity of the test set were 0.927, 0.931 and 0.921, respectively, which confirmed the effectiveness and feasibility of the subtraction new modal features in split kidney function prediction.

[0062] Those skilled in the art can understand that the module schematic diagram is only an example of the cortical medulla enhancement phase CTU kidney cortex segmentation and split kidney function prediction based on artificial intelligence neural network, and does not constitute a limitation on the terminal equipment of the method, and can include more or fewer components than the schematic diagram, or combine certain components, or use different components. For example, the method can also include input and output devices, network access devices, buses, etc.

[0063] The method can also include computing devices such as computers, notebooks, palmtop computers, and cloud servers, and include but not limited to processing modules, storage modules, and computer programs stored in the storage module and executable on the processing module, such as the segmentation program described above. The processing module, when executing the computer program, implements the steps in each of the above kidney cortex segmentation and modeling method embodiments, such as the steps shown in Figure 1 .

[0064] In addition, the present application also proposes an electronic device including a storage module and a processing module, which can implement the steps in the above kidney cortex segmentation and modeling method when executing the computer program, that is, implement the steps in any one of the above technical solutions of the kidney cortex segmentation method. The electronic device can be part of the device integrated in the method, or a local terminal device, and can also be part of a cloud server.

[0065] The processing module can be a central processing unit (CPU), and can also be other general-purpose processors, graphics processing units (GPUs), digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processing module is the control center of the method, and connects various parts of the system through various interfaces and lines.

[0066] The storage module can be used to store the computer program and / or the module, and the processing module realizes various functions of the method by running or executing the computer program and / or the module stored in the storage module, and calling the data stored in the storage module. The storage module can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required by a function, etc. In addition, the storage module can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, an intelligent memory card, a secure digital card, a flash memory card, at least one disk storage device, a flash memory device or other volatile solid-state storage device.

[0067] Exemplarily, the computer program can be divided into one or more modules / units, which are stored in the storage module and executed by the processing module. The one or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the method.

[0068] In addition, an embodiment of the present application provides a readable storage medium storing a computer program, which can realize the steps in the renal cortex segmentation method when executed by the processing module.

[0069] The modules of the kidney cortex segmentation method described in the present application, if realized in the form of software function units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer-readable storage medium. The computer program can realize the steps of each method embodiment when executed by a processing module.

[0070] The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0071] It should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description manner of the specification is only for the sake of clarity, and the person skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by the person skilled in the art.

[0072] The person skilled in the art will realize that the embodiments described herein are to help the reader to understand the principles of the present application, and should be understood as the protection scope of the present application is not limited to such specific descriptions and embodiments. The person skilled in the art can make various specific modifications and combinations that do not deviate from the essence of the present application according to the technical inspirations disclosed in the present application, and these modifications and combinations are still within the protection scope of the present application.

Claims

1. A method for predicting unilateral renal function grading based on CT subtraction imaging, characterized in that: The following steps are involved: Step 1: Extract the patient's cortical and medullary enhancement phase CTU images to obtain plain scan CTU images and enhanced CTU images; Step 2: Align the plain scan CTU image and the enhanced CTU image based on physical space resampling and affine rigid registration; Step 3: Use the neural network model to automatically segment the bilateral renal cortex and further infer to obtain the renal cortex mask; Step 4: Based on the registered enhanced image and the plain scan image output in step 2, pixel-by-pixel subtraction is performed to construct the subtraction imaging modality; Step 5: Extract the radiomic features of the subtraction imaging modality to obtain the radiomic features of the unilateral kidney; Step 6: Predict the renal function based on the renal function prediction model.

2. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 1, wherein: The specific implementation of step 2 includes the following sub-steps: Step 2.1: Using the enhanced CTU image as the reference sequence, create a resampler based on the size, resolution, origin position, and orientation matrix of the enhanced CTU to resample the CT plain scan image sequence to the same coordinate space as the enhanced CT. Step 2.2: Use linear interpolation to resample the CT grayscale image. Set the transformation to unity and resample the image to the new spatial frame. For pixels that do not have a corresponding original image position in the new space, assign a default pixel value of 0. Step 2.3: Unify the image orientation to "left-back-top" and initialize an affine transformation based on the geometric center; Step 2.4: Set the registration strategy and multi-resolution strategy, perform the registration, and obtain the final transformation matrix; then map the plain CT image to the spatial frame of the enhanced CT according to this transformation.

3. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 2, wherein: In step 2.4, the registration strategy uses MattesMutualInformation as the similarity metric function, the optimizer is set to GradientDescent, and the learning rate, maximum number of iterations, convergence threshold, and convergence window are set. The metric_record callback function is called at each iteration to record the current metric value to facilitate analysis of the convergence process.

4. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 1, characterized in that: In step 3, the bilateral renal cortical structures are first annotated layer by layer for the input image data to generate a corresponding mask file; then the non-zero area data is cropped; followed by data resampling and normalization. The resampling is used to normalize the intervals of all images to the median interval of the data set. The original data is resampled using third-order spline interpolation, while the mask data uses the nearest neighbor interpolation method; finally, the neural network model is input for image segmentation, and further reasoning is performed to obtain the renal cortex mask; The neural network model consists of 7 different stages, each of which is responsible for extracting features at different levels, gradually from low-level features to high-level features; The number of feature channels in each stage is [32, 64, 128, 256, 320, 320,320]; 3D convolution is used in each stage to capture spatial information, and the convolution kernel sizes of each stage are: Stage 1: [1, 3,3]; Stage 2: [1, 3, 3]; Stage 3-7: [3, 3, 3]; the stride is set to: Stage 1: [1, 1, 1]; Stage 2: [1, 2, 2]; Stage 3: [1, 2, 2]; Stage 4: [2, 2, 2]; Stage 5: [2, 2,2]; Stage 6: [1, 2, 2]; Stage 7: [1, 2, 2]; each stage of the encoder and decoder has 2 convolution layers [2, 2, 2, 2, 2, 2, 2]; Do not use the Dropout layer; Specify the nonlinear activation function used as torch.nn.LeakyReLU, and the nonlinear activation function is set to be calculated in situ.

5. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 4, characterized in that: In step 3, the probability map generated by the neural network model is converted into a binary mask through a threshold, and is used to remove artifacts and noise smaller than the set threshold through dilation and erosion operations; the connected components in the segmentation result are analyzed, and the connected areas smaller than the set threshold are removed; finally, the segmentation result is smoothed to output an accurate probability map.

6. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 1, characterized in that: In step 3, the neural network model is a trained model; During the training process, the loss function used Combined with Dice loss and cross entropy loss ; in, is the Dice loss function; K is the set of categories; |K| is the number of categories; k∈K: the kth category currently being calculated; I is the set of positions of all pixels or voxels in the image; i∈I: the position of the pixel or voxel currently being calculated; u is the Softmax probability output, and v is the hard-coded ground truth; is the model's predicted value for the i-th pixel in the k-th class; is the value of the i-th pixel in the k-th class in the true label; is the intersection of the prediction and the true label on the kth category; is the union of the total number of predictions and actual labels on the kth class; is the cross entropy loss value; Y is the true label; P is the predicted probability of the model output; Learning rate adjustment strategy during training: using Adam optimizer, updated weight parameters ; Represents the convolution weight parameter before updating; is the current learning rate; represents the bias-corrected first-order moment estimate; is the second-order moment estimate after bias correction; ε is a small constant to prevent the denominator from being zero and maintain numerical stability; Stop training when the epoch limit is reached, or when the exponential moving average loss of the validation set decreases by no more than 5e-3 within 60 training epochs, or when the learning rate is reduced to 1e-6; During training, a variety of data augmentation techniques are used, including random rotation, random scaling, random elastic transformation, gamma correction, and mirroring, to improve the generalization ability of the model; During the training process, in order to increase the stability of the network, patch sampling ensures that more than 1 / 3 of the pixels in a batch are foreground pixels.

7. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 1, characterized in that: In step 4, the enhanced CT image after registration is first subtracted from the corresponding plain scan CT image voxel by voxel. Then, to remove minor differences and artifacts, a tolerance threshold is introduced. For all subtracted voxels, if the difference in enhancement value is less than the threshold, it is considered to have no significant enhancement and is regarded as background.

8. The method for predicting unilateral renal function grading based on CT subtraction imaging according to claim 1, characterized in that: In step 6, the renal function prediction model adopts the ElasticNet regression model, and the reference objective function is: Among them, y i is the actual observation value of the i-th sample; x ij is the jth input feature value of the i-th sample; n is the total number of samples; p is the total number of features; β0 is the intercept of the model; β j is the weight coefficient of the j-th feature, i.e., the model parameter; is the regularization strength coefficient, is the weight ratio of the L1 norm and L2 norm regularization terms; The renal function prediction model is a trained model; During the training process, the patients' matched renal parenchymal dynamic phenomena 99mTc-DTPA examination results were collected simultaneously, and the GFR was calculated based on the 99mTc-DTPA renal dynamic imaging using the Gates analysis method: During the training process, a 5-fold cross-validation strategy was used to ensure the robustness and generalization of the model.

9. A grading system for predicting unilateral renal function based on CT subtraction imaging, characterized in that: Includes the following modules: Image acquisition and preprocessing module, used to extract the patient's cortical and medullary enhancement phase CTU images, and obtain plain scan CTU images and enhanced CTU images; Multimodal CTU image registration module based on physical space resampling and affine rigid registration, used to align plain scan CTU images and enhance CTU images based on physical space resampling and affine rigid registration; The neural network-based segmentation module is used to automatically segment the bilateral renal cortex using a neural network model and further infer the renal cortex mask; The subtraction processing module is used to construct the subtraction imaging modality by pixel-by-pixel subtraction based on the registered enhanced image and the plain scan image; Feature extraction module, used to extract radiomic features of subtraction imaging modalities and obtain unilateral renal radiomic features; A renal function prediction model is constructed to predict renal function based on the renal function prediction model.

10. A product for predicting unilateral renal function grading based on CT subtraction imaging, comprising computer program instructions, characterized in that: When the computer program instructions are executed on a computer, the computer is caused to execute the method for predicting unilateral renal function grading based on CT subtraction imaging according to any one of claims 1 to 8.

Images