Image evaluation method and device for activity of common solid tumors in abdomen of child and medium

Abstract

The invention provides a child abdominal common solid tumor activity image evaluation method and device and a medium, and the method comprises the steps: capturing the change information of a focus on a vein phase image after plain scanning and enhancement through a 3D convolutional neural network model, carrying out the intelligent silhouette, and generating an enhancement region and the enhancement degree of the focus after chemotherapy, thereby achieving the evaluation of the activity of a child abdominal common solid tumor. And finally, integrating all the information to perform intelligent diagnosis, judging the tumor activity, and forming an artificial intelligence auxiliary activity evaluation scheme after chemotherapy of the abdominal solid tumor. According to the method, the automatic enhancement degree quantitative evaluation of the hepatoblastoma, the nephroblastoma and the neuroblastoma of the children is realized through automatic tumor segmentation and registration subtraction image calculation by utilizing the plain scanning and enhanced CT image, so that the focus characteristics are extracted from multiple angles, the image information is fully mined, and the accuracy of the diagnosis is improved. Accuracy and sensitivity of activity discrimination after chemotherapy of the solid abdominal tumor of children can be obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, a device and a computer-readable storage medium for evaluating the activity of common solid tumors in the abdomen of children. Background Art

[0002] Solid tumors in the pediatric body account for about 40% of all pediatric malignant tumors, and are most likely to occur in the abdomen. Neuroblastoma, Wilms tumor, and hepatoblastoma are the most common, which seriously endanger the health of children in my country. Surgical resection combined with neoadjuvant chemotherapy and postoperative chemoradiotherapy has become the standard treatment for common solid tumors in the pediatric abdomen. Accurate imaging evaluation is an important method to judge the efficacy of neoadjuvant chemotherapy, determine the timing of surgery, and monitor postoperative recurrence.

[0003] Long-term experience from multidisciplinary comprehensive diagnosis and treatment (MDT) and clinical drug trials shows that after neoadjuvant chemotherapy, tumor reduction is often asymmetric, and some only show a reduction in the range or degree of enhancement in the tumor in the early stages of chemotherapy. In addition, common solid tumors in the abdomen of children are usually large in size, irregular in shape, and easy to push, wrap around, and invade surrounding organs and large blood vessels, which poses a challenge to accurate imaging evaluation.

[0004] At present, the double-diameter method or the three-diameter method is still widely used in clinical practice to assess tumor size, which has poor accuracy. The three-dimensional quantitative technology of layer-by-layer delineation and measurement of tumor volume can obtain the actual size of the tumor and avoid the error caused by asymmetric reduction. However, manual delineation of the tumor is time-consuming and laborious. For large tumors, manual measurement time can reach more than 1 hour. Moreover, some tumors have irregular boundaries and low contrast with surrounding structures. Pure manual delineation depends on the judgment of the clinical experience of the measurer. The results between measurers have large variability and poor repeatability, which makes it difficult to promote and apply. In addition, the evaluation of tumor function and activity is also an important aspect of tumor imaging evaluation. At present, the most commonly used clinical methods are CT / MRI enhanced examination or MRI diffusion-weighted imaging. CT / MRI enhanced examination indirectly reflects the degree of tumor activity by observing the current blood supply status of the tumor, and can reflect the tumor efficacy earlier before the tumor volume changes. However, for pediatric solid tumors with large anatomical spans, the evaluation of the range and degree of enhancement is more experience-dependent. When the volume change is not obvious, young doctors and grassroots doctors may even come to the opposite conclusion from senior experts, which seriously affects the choice of subsequent treatment options and prognosis of children. Therefore, there is an urgent need for intelligent assessment tools with high accuracy and timeliness that can take into account both tumor volume measurement and functional analysis, so as to achieve accurate imaging assessment of common solid tumors in the abdomen of children. Summary of the invention

[0005] The present invention aims to provide a method, device and computer-readable storage medium for active imaging assessment of common solid tumors in the abdomen of children, which overcome the above problems or at least partially solve the above problems.

[0006] To achieve the above object, the technical solution of the present invention is specifically implemented as follows:

[0007] One aspect of the present invention provides a method for evaluating the activity of common solid tumors in the abdomen of children by imaging, comprising:

[0008] Determining a tumor segmentation model, wherein the tumor segmentation model is a UNETR model;

[0009] Resample the input image according to a preset standard spacing;

[0010] Block sampling is performed on the resampled data according to a preset block size;

[0011] Using test-time augmentation techniques, training samples are randomly rotated and augmented;

[0012] Using the training set to train the tumor segmentation model with the segmentation and annotation results of the tumor area annotated by the physician as a target, to obtain a trained tumor segmentation model;

[0013] Using the validation set to validate and optimize parameters of the trained tumor segmentation model;

[0014] Using the test set to evaluate the trained tumor segmentation model;

[0015] Segmenting the image to be segmented using the trained tumor segmentation model to obtain a tumor segmentation result, and visualizing the tumor segmentation result;

[0016] Acquire a child CT image, and remove artifacts from the child CT image;

[0017] To register the children's CT plain scan images after removing artifacts with the arterial phase images and venous phase images;

[0018] Directly mapping the tumor area in the tumor segmentation result to the plain scan and arterial phase images after registration transformation, respectively calculating the subtraction images of the arterial phase, venous phase and plain scan phase to obtain the arterial and venous absolute enhancement value images of the tumor, and dividing the absolute enhancement value image by the plain scan phase image to obtain a normalized enhancement image;

[0019] The absolute enhancement value image and the normalized enhancement image are used to evaluate the enhancement degree.

[0020] Optionally, the tumor segmentation model is trained using the training set with the segmentation and annotation results of the tumor area annotated by the physician as a target, and the trained model includes:

[0021] The back propagation algorithm and the stochastic gradient descent method are used to iteratively update the weight of each layer through back propagation according to the size of the loss function between the predicted segmentation result and the labeling result of the forward propagation. When the loss function between the predicted segmentation result and the labeling result of the model tends to converge, the training of the model is stopped to obtain the trained model.

[0022] Optionally, the using the validation set to validate and optimize parameters of the trained tumor segmentation model includes:

[0023] By monitoring the Dice and IoU coefficients, the best model parameters and hyperparameter configurations are selected.

[0024] Optionally, the preset standard spacing is 1.0×1.0×1.0 mm; and the preset block size is 128×128×128 voxels.

[0025] Optionally, the acquiring of a child CT image and performing artifact removal on the child CT image includes:

[0026] Decomposing the child CT image into different frequency sub-bands using wavelet transform;

[0027] Detecting textures manifesting as high-frequency noise or anomalies in the decomposed subbands and thresholding the detected artifacts;

[0028] The sub-bands after threshold processing are reconstructed back to the image space through inverse wavelet transform to obtain the CT image after artifact removal.

[0029] Optionally, the registering of the artifact-removed pediatric CT plain scan image and the arterial phase image and the venous phase image comprises:

[0030] The tumor segmentation result and the high-density bones in the abdomen are used to perform auxiliary feature point registration, wherein the registration evaluation standard adopts mutual information MI.

[0031] Optionally, the using the tumor segmentation result and the abdominal high-density bone to perform auxiliary feature point registration includes:

[0032] The artifact-free pediatric CT plain scan images were normalized before registration using the Z score;

[0033] The high-density bone is segmented by the Otsu adaptive threshold segmentation method, and the cavity is filled by a closing operation, and the edge points and corner points of the obtained tumor and high-density bone area are extracted respectively, wherein the edge points are detected by the Canny edge detection algorithm and the points are obtained by selecting points at equal intervals on the edge, and the corner points are obtained by the Harris corner point detection algorithm;

[0034] Using a nearest neighbor matching algorithm to find corresponding feature point pairs in the artifact-removed pediatric CT plain scan image, the arterial phase image, and the venous phase image;

[0035] Use affine transformation for registration and estimate the transformation parameters by minimizing the mean square error of matching point pairs;

[0036] The similarity measure is defined using mutual information and feature point matching error;

[0037] Maximize the similarity measure via gradient descent;

[0038] Iteration is performed until the similarity measure no longer increases significantly and the feature point matching error is less than a preset value, then the iteration is stopped and the registration is completed.

[0039] Optionally, the using the absolute enhancement value image and the normalized enhancement image to perform enhancement degree assessment includes:

[0040] Calculating enhancement distribution information of the patient's tumor area according to the normalized enhanced image;

[0041] The proportion of the part of the tumor with enhancement greater than the set threshold to the total tumor volume was calculated.

[0042] Another aspect of the present invention provides a device for evaluating the activity of common solid tumors in the abdomen of children, comprising: a processor, a memory;

[0043] The memory is used to store computer programs;

[0044] The processor is used to execute the above-mentioned method for evaluating the activity of common solid tumors in the abdomen of children by calling the computer program.

[0045] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the above-mentioned method for imaging the activity of common solid tumors in the abdomen of children.

[0046] It can be seen that the method, device and computer-readable storage medium for evaluating the activity of common solid tumors in the abdomen of children provided by the present invention use a 3D convolutional neural network model to capture the change information of the lesion on the plain scan and enhanced venous phase images, and intelligently generate the enhanced area and degree of the lesion after chemotherapy by silhouette, and finally integrate all the information to perform intelligent diagnosis and judge the tumor activity, thereby forming an artificial intelligence-assisted activity evaluation scheme for abdominal solid tumors after chemotherapy; the present invention utilizes plain scan and enhanced CT images, and realizes the automatic enhancement degree quantitative evaluation of childhood hepatoblastoma, nephroblastoma and neuroblastoma through automatic tumor segmentation and registration subtraction image calculation, thereby extracting lesion characteristics from multiple angles, fully mining image information, and significantly improving the accuracy and sensitivity of activity discrimination of childhood abdominal solid tumors after chemotherapy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0048] Figure 1 A flow chart of a method for evaluating the activity of common solid tumors in the abdomen of children provided by an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of the UNETR network structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0051] In order to solve the problems of the prior art, the present invention proposes a method, device and computer-readable storage medium for evaluating the activity of common solid tumors in the abdomen of children based on deep learning. The present invention uses plain scan and enhanced CT images, and realizes automatic quantitative evaluation of the degree of enhancement of hepatoblastoma, nephroblastoma and neuroblastoma in children through automatic tumor segmentation and registration subtraction image calculation.

[0052] Figure 1 A flowchart showing a method for evaluating the activity of common solid tumors in the abdomen of children provided by an embodiment of the present invention is shown. Figure 1The embodiment of the present invention provides a method for evaluating the activity of common solid tumors in the abdomen of children, comprising:

[0053] 1. Build a gold standard database and a tumor segmentation model, including:

[0054] Specifically, the tumor segmentation model of the present invention is designed as a Transformer-based UNETR network. In the field of medical image segmentation, deep learning technology has made significant progress. UNETR is a Transformer-based segmentation model that introduces the self-attention mechanism into the image segmentation task in order to obtain better performance and generalization ability. The present invention trains a universal pediatric solid tumor segmentation model for different cancers (NB / WT / HB) on a data set with tumor-labeled Mask information to achieve tumor segmentation.

[0055] S1, determining a tumor segmentation model, wherein the tumor segmentation model is a UNETR model.

[0056] Specifically, the present invention selects the UNETR model, such as Figure 2 As shown in the figure, this is an advanced model that combines U-Net and Transformer. It not only retains the encoder-decoder structure of U-Net to capture the contextual information of the image, but also enhances the ability to capture global dependencies through the self-attention mechanism of Transformer, thus better preserving detailed information.

[0057] S2, resampling the input image according to a preset standard spacing;

[0058] S3, performing block sampling on the resampled data according to a preset block size.

[0059] As an optional implementation of the embodiment of the present invention, the preset standard spacing is 1.0×1.0×1.0 mm; the preset block size is 128×128×128 voxels.

[0060] Specifically, this step only wants to resample the image and unify the voxels:

[0061] In the specific implementation, resampling the input image to unify the voxel spacing is a key step to ensure consistency between different images. Especially in the UNETR model using the self-attention mechanism, a unified voxel spacing is crucial to maintain the consistency of the model's performance on various images. The present invention selects the median voxel spacing in the dataset, i.e., 1.0×1.0×1.0 mm, as the standard spacing for resampling.

[0062] After resampling, block sampling is further implemented to obtain a more uniform data resolution. Since the UNETR model incorporates the self-attention mechanism, it is more sensitive to the resolution of the input data and the size of the block. Therefore, the present invention designs the block size and sampling strategy to maximize the performance of the model. After optimization, the present invention determines that the block size is unified as 128×128×128 voxels, which can ensure sufficient local feature capture and meet the model's demand for global context information.

[0063] S4, uses test time augmentation techniques to randomly rotate and expand the training samples.

[0064] Specifically, this step performs test time enhancement and model generalization:

[0065] In specific implementation, the present invention uses test time enhancement technology to randomly rotate and expand the training samples to enhance the generalization ability of the model for different cancers. Since the UNETR model has better global feature capture capabilities, the robustness of the model can be further improved through data enhancement strategies such as color transformation and noise addition.

[0066] S5, using the training set to train the tumor segmentation model with the segmentation and annotation results of the tumor area annotated by the physician as a target, to obtain a trained tumor segmentation model.

[0067] As an optional implementation of an embodiment of the present invention, a tumor segmentation model is trained using a training set with the segmentation and annotation results of the tumor area annotated by a physician as the target, and the trained model includes: using a backpropagation algorithm and a stochastic gradient descent method, and performing backpropagation iteratively to update the weights of each layer according to the size of the loss function between the predicted segmentation result and the annotation result of the forward propagation, until the loss function between the predicted segmentation result and the annotation result of the model tends to converge, then stopping the training of the model to obtain the trained model.

[0068] Specifically, the present invention uses a training set to train the UNETR model with the segmentation and annotation results of the tumor area annotated by the physician as the target, adopts the back propagation algorithm and the stochastic gradient descent method, and performs back propagation iteratively to update the weight of each layer according to the size of the loss function between the predicted segmentation result and the annotation result of the forward propagation, until the loss function between the predicted segmentation result and the annotation result of the model tends to converge, stops training the model, and obtains the trained model.

[0069] S6, use the validation set to verify and tune the parameters of the trained tumor segmentation model.

[0070] As an optional implementation of the embodiment of the present invention, using the validation set to validate and tune parameters of the trained tumor segmentation model includes: selecting optimal model parameters and hyperparameter configurations by monitoring Dice and IoU coefficients.

[0071] Specifically, the present invention performs model verification and parameter tuning on the validation set. By monitoring the Dice and IoU coefficients, the optimal model parameters and hyperparameter configurations are selected. The loss function of the present invention is a combination of Dice loss and cross entropy loss.

[0072] The Dice loss is shown below:

[0073]

[0074] Among them, p i represents the predicted value, g i represents a real value, and the sum is performed over all elements.

[0075] The cross entropy loss is shown below:

[0076] L cE =-∑ i g i log(p i )

[0077] Among them, g i is the one-hot encoding of the true label, p i is the predicted probability distribution.

[0078] The total loss function is as follows:

[0079] L loss =L Dice +λL CE

[0080] Among them, λ is a balancing coefficient used to adjust the weight between the two losses.

[0081] S7, Evaluation of the trained tumor segmentation model using the test set.

[0082] Specifically, the trained UNETR segmentation model is evaluated on the test set. The Dice coefficient is a key indicator for measuring the accuracy of tumor segmentation tasks. It evaluates the accuracy and consistency of segmentation results by calculating the overlap between the predicted segmentation and the true segmentation annotated by senior radiologists. The Dice coefficient score ranges from 0 to 1. The closer the score is to 1, the higher the overlap between the predicted segmentation and the standard annotation, and the stronger the similarity. Generally speaking, a Dice coefficient exceeding 0.5 means that the segmentation results have a high degree of overlap, which reflects a better segmentation performance.

[0083]

[0084] Among them, A represents the binary segmentation result predicted by the model, B represents the binary segmentation result of the gold standard, and |·| represents the number of pixels.

[0085] Second, automatically extract the tumor area from the image, including:

[0086] S8, using the trained tumor segmentation model to segment the image to be segmented, obtaining a tumor segmentation result, and visualizing the tumor segmentation result.

[0087] 3. Registration of plain scan and arterial phase images to venous phase images, including:

[0088] S9, obtaining a child CT image, and performing artifact removal on the child CT image.

[0089] As an optional implementation of an embodiment of the present invention, obtaining a child CT image and removing artifacts from the child CT image includes: using wavelet transform to decompose the child CT image into different frequency sub-bands; detecting textures that appear as high-frequency noise or abnormalities in the decomposed sub-bands, and performing threshold processing on the detected artifacts; and reconstructing the sub-bands after threshold processing back to the image space through inverse wavelet transform to obtain a CT image after artifacts are removed.

[0090] Specifically, in pediatric CT scanning, since children have smaller bodies and their organ shapes and sizes are different from those of adults, and children may move during the scanning process, artifacts may be generated in the obtained CT images. Traditional artifact removal methods focus on the reconstruction process from raw projection data to CT images, while the present invention only has reconstructed CT images, so the following artifact removal method based on pediatric CT images is proposed.

[0091] 1) Image decomposition

[0092] Use wavelet transform to decompose the image into different frequency subbands:

[0093] {c A ,c D1 ,c D2 ,…,c DB}=W(I)

[0094] Where I is the CT image containing artifacts, is the wavelet transform, c A is the approximation coefficient, c D1 ,c D2 ,…,c DN is the detail factor.

[0095] 2) Artifact detection and threshold processing

[0096] In the decomposed sub-bands, textures that appear as high-frequency noise or abnormalities are detected, and the detected artifacts are thresholded to reduce their intensity or completely removed. At the same time, in order to retain the image details while removing high-frequency artifacts, the threshold part uses an adaptive threshold for local adjustment. The threshold is automatically adjusted according to the local statistical characteristics (standard deviation) of the image, thereby removing artifacts while retaining useful details.

[0097] τ(i,j)=k·σ(i,j)

[0098] c′ D (i,j)=sign(c D (i,j))·max(|c D (i,j)|-τ(i,j),0)

[0099] where c D (i, j) is the wavelet detail coefficient at position (i, j), σ(i, j) is the local standard deviation at that position, k is a constant, τ(i, j) is the calculated adaptive threshold at that point, sign(c D (i,j)) is c D The sign of (i,j).

[0100] 3) Image reconstruction

[0101] The subbands processed by adaptive threshold are reconstructed back to the image space through inverse wavelet transform:

[0102]

[0103] Where I′ is the CT image after artifact removal.

[0104] S10, registering the child CT plain scan image after artifact removal with the arterial phase image and the venous phase image.

[0105] As an optional implementation of an embodiment of the present invention, registering the artifact-removed pediatric CT plain scan image with the arterial phase image and the venous phase image includes: using the tumor segmentation result and the abdominal high-density bone to perform auxiliary feature point registration, wherein the registration evaluation criterion adopts mutual information MI. Among them, using the tumor segmentation results and the high-density bones in the abdomen for auxiliary feature point registration includes: standardizing the children's CT plain scan images after artifact removal before registration through the Z score; segmenting the high-density bones through the Otsu adaptive threshold segmentation method, and filling the holes through the closing operation, and extracting edge points and corner points from the obtained tumor and high-density bone areas respectively, wherein the edge points are detected by the Canny edge detection algorithm and the points are selected at equal intervals on the edge, and the corner points are obtained by the Harris corner detection algorithm; the nearest neighbor matching algorithm is used to find the corresponding feature point pairs in the children's CT plain scan images, arterial phase images and venous phase images after artifact removal; affine transformation is used for registration, and the transformation parameters are estimated by minimizing the mean square error of the matching point pairs; the similarity measure is defined by mutual information and feature point matching error; the similarity measure is maximized by the gradient descent method; it is iterated until the similarity measure no longer increases significantly and the feature point matching error is less than the preset value, and the iteration is stopped, and the registration is completed.

[0106] Specifically, the data used in the present invention are plain scan and enhanced images of pediatric abdominal CT. Due to the different acquisition times of different images, there is a mismatch between the sequences. In order to conduct subsequent enhancement assessment, it is necessary to first align the plain scan and arterial and venous phase images. Clinically, radiologists generally believe that the tumor boundary of the venous phase image is clearer, so the venous phase image is used as the reference image in the registration process, and the plain scan and arterial phase images are registered to the venous phase image.

[0107] Considering that children with abdominal solid tumors often have abdominal space-occupying lesions, it is difficult to assist in organ registration. Therefore, the present invention uses the above tumor segmentation results and abdominal high-density bones to assist in feature point registration. At the same time, considering that the grayscale distribution of plain scan, arterial, and venous phase images is different due to contrast agents, the registration evaluation standard is selected through mutual information (MI). MI measures the similarity between two images and does not require the images to have the same grayscale distribution. Therefore, it is more suitable for the present invention. The specific registration method is as follows:

[0108] 1) Image standardization

[0109] Normalize the images before registration using Z-scores:

[0110]

[0111] Where I′ is the standardized image, min(I) and max(I) are the minimum and maximum values ​​in the image respectively.

[0112] 2) Feature point extraction

[0113] The tumor area is obtained through the automatic segmentation model. The high-density bone is automatically segmented from the image through the Otsu adaptive threshold segmentation method:

[0114] Bone=(I>τ B )

[0115] where τ B is the image adaptive threshold.

[0116] The bone area obtained by threshold segmentation usually produces discrete small points due to the low-density area in the bone, and the holes are automatically filled by closing operation.

[0117] Bone c =C(Bone)

[0118] Where C is a closing operation.

[0119] Edge points and corner points (bone ridge points) are extracted from the obtained tumor and high-density bone areas respectively, and edge detection is performed using the Canny edge detection algorithm. Edge points are obtained by selecting points at equal intervals on the edge, and corner points are obtained using the Harris corner point detection algorithm.

[0120] E Tumor =Canny(Tumor)

[0121] E Bone =Canny(Bone)

[0122] C Tumor =Harris(Tumor)

[0123] C Bone =Harris(Bone)

[0124] Where E Tumor , E Bone , C Tumor , c Bone are the edge points and corner points of the tumor and bone respectively.

[0125] 3) Feature point matching

[0126] Use the nearest neighbor matching algorithm to find the corresponding feature point pairs in the plain scan, arterial phase image, and venous phase image:

[0127]

[0128] Where p is a feature point in the plain scan or arterial phase image, q is the corresponding feature point in the venous phase image, d(p,q) is the Euclidean distance between points p and q, and Q is the feature point set of the venous phase image.

[0129] 4) Transformation model estimation

[0130] Since the overall shapes of the phases are similar, affine transformation is used for registration, and the transformation parameters are estimated by minimizing the mean square error of the matching point pairs:

[0131] T(x;θ)=Ax+t

[0132] Where T is the transformation function, x is the original coordinate, θ is the transformation parameter, A is a 2×2 matrix, and t is the translation vector.

[0133] 5) Similarity measurement

[0134] The similarity measure is defined using mutual information and feature point matching error:

[0135] S(θ)=αMI(T(I 1 ; θ),I 2 )-βD(M,T(P 1 ;θ))

[0136]

[0137] Where MI is the mutual information, D is the feature point matching error, and α and β are the weight coefficients of the two.

[0138] 6) Optimize registration

[0139] Maximize the similarity measure via gradient descent:

[0140]

[0141] in is the gradient of the similarity measure with respect to the transformation parameter θ, and λ is the learning rate.

[0142] 7) Iteration

[0143] Repeat steps 3)-6) until the similarity measure no longer increases significantly and the feature point matching error is less than the preset value, then stop the iteration and the registration is completed.

[0144] 4. Calculate the normalized tumor enhancement image based on the subtraction method, including:

[0145] S11, directly mapping the tumor area in the tumor segmentation result to the plain scan and arterial phase images after the registration transformation, calculating the subtraction images of the arterial phase, venous phase and plain scan phase respectively, obtaining the arterial and venous absolute enhancement value images of the tumor, and dividing the absolute enhancement value image by the plain scan phase image to obtain the normalized enhancement image.

[0146] Specifically, the tumor area obtained by the above tumor segmentation model can be directly mapped to the plain scan and arterial phase images after registration transformation. The subtraction images of the arterial phase, venous phase and plain scan phase are calculated respectively to obtain the absolute enhancement value images of the artery and vein of the tumor. In order to reduce the influence of individual differences between different patients, the obtained absolute enhancement value image is divided by the plain scan phase image to obtain a normalized enhancement image.

[0147] I sub =I post -I pre

[0148]

[0149] Among them I post For arterial or venous phase images, I pre is the image during the flat scan period, I sub and I′ sub They are the calculated absolute enhancement value image and normalized enhancement image respectively.

[0150] 5. Quantify tumor area enhancement information, including:

[0151] S12, using the absolute enhancement value image and the normalized enhancement image to evaluate the enhancement degree.

[0152] As an optional implementation of an embodiment of the present invention, using an absolute enhancement value image and a normalized enhancement image to evaluate the degree of enhancement includes: calculating the enhancement distribution information of the patient's tumor area based on the normalized enhancement image; calculating the proportion of the part of the tumor with enhancement greater than a set threshold in the total tumor volume.

[0153] Specifically, the present invention first calculates the enhancement distribution information of the patient's tumor area based on the calculated normalized enhanced image:

[0154]

[0155] Among them, H sub (x) is used to calculate the histogram, p(x i ) is I′ sub (x i ) in x i The probability density at I′ sub (x i ) is in x i The pixel value at .

[0156] The proportion of the part of the tumor with enhancement greater than the set threshold to the total tumor volume was calculated, where the threshold was set in increments of 0.2 to 5.

[0157]

[0158] Where V sub is the tumor volume, V enhenced The degree of reinforcement is greater than the set threshold T threshold VolumeRatio is the volume ratio that meets the requirements obtained by final calculation.

[0159] It can be seen that the method for evaluating the activity of common solid tumors in the abdomen of children provided by the embodiment of the present invention applies deep learning convolutional neural network technology to the feature learning of abdominal enhanced CT medical image data of solid tumors in the abdomen of children, and performs model optimization; analyzes the changes in the degree of enhancement in the venous phase after plain scan and enhancement, thereby improving the ability of model activity evaluation.

[0160] Furthermore, the present invention can be used to extract two-dimensional and high-order features from the input abdominal enhanced CT image frames through a neural network to improve the discrimination ability of the model; at the same time, a corresponding computer-aided diagnosis system can be used to assist clinical doctors in making high-level diagnoses, allowing grassroots doctors and young doctors to obtain diagnostic capabilities that are homogeneous with senior experts, thereby effectively improving the level of imaging assessment of abdominal solid tumors in children in my country.

[0161] The present invention also provides a device for evaluating the activity of common solid tumors in the abdomen of children, comprising: a processor, a memory;

[0162] Memory for storing computer programs;

[0163] The processor is used to execute the above-mentioned imaging method for evaluating the activity of common solid tumors in the abdomen of children by calling a computer program.

[0164] It can be seen that the activity image assessment device for common solid tumors in the abdomen of children provided by the embodiment of the present invention applies deep learning convolutional neural network technology to the feature learning of abdominal enhanced CT medical image data of solid tumors in the abdomen of children, and performs model optimization; analyzes the changes in the degree of enhancement in the venous phase after plain scan and enhancement, thereby improving the ability of model activity assessment.

[0165] The present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the above-mentioned method for active imaging assessment of common solid tumors in the abdomen of children.

[0166] It can be seen that through the computer-readable storage medium provided by the embodiment of the present invention, the deep learning convolutional neural network technology is applied to the feature learning of abdominal enhanced CT medical image data of children's abdominal solid tumors, and the model is optimized; the changes in the degree of enhancement in the venous phase after plain scanning and enhancement are analyzed, thereby improving the ability to evaluate the activity of the model.

[0167] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for evaluating the activity of common solid tumors in the abdomen of children, characterized in that: include: Determining a tumor segmentation model, wherein the tumor segmentation model is a UNETR model; Resample the input image according to a preset standard spacing; Block sampling is performed on the resampled data according to a preset block size; Using test-time augmentation techniques, training samples are randomly rotated and augmented; Using the training set to train the tumor segmentation model with the segmentation and annotation results of the tumor area annotated by the physician as a target, to obtain a trained tumor segmentation model; Using the validation set to validate and optimize parameters of the trained tumor segmentation model; Using the test set to evaluate the trained tumor segmentation model; Segmenting the image to be segmented using the trained tumor segmentation model to obtain a tumor segmentation result, and visualizing the tumor segmentation result; Acquire a child CT image, and remove artifacts from the child CT image; To register the artifact-free pediatric CT plain scan images with the arterial phase images and venous phase images; Directly mapping the tumor area in the tumor segmentation result to the plain scan and arterial phase images after registration transformation, respectively calculating the subtraction images of the arterial phase, venous phase and plain scan phase to obtain the arterial and venous absolute enhancement value images of the tumor, and dividing the absolute enhancement value image by the plain scan phase image to obtain a normalized enhancement image; The absolute enhancement value image and the normalized enhancement image are used to evaluate the enhancement degree.

2. The method according to claim 1, characterized in that The tumor segmentation model is trained using the training set with the segmentation and annotation results of the tumor area annotated by the physician as the target, and the trained model includes: The back propagation algorithm and the stochastic gradient descent method are used to iteratively update the weight of each layer through back propagation according to the size of the loss function between the predicted segmentation result and the labeling result of the forward propagation. When the loss function between the predicted segmentation result and the labeling result of the model tends to converge, the training of the model is stopped to obtain the trained model.

3. The method according to claim 1, characterized in that: The using the validation set to validate and tune parameters of the trained tumor segmentation model includes: By monitoring the Dice and IoU coefficients, the best model parameters and hyperparameter configurations are selected.

4. The method according to claim 1, characterized in that The preset standard spacing is 1.0×1.0×1.0 mm; the preset block size is 128×128×128 voxels.

5. The method according to claim 1, characterized in that The acquiring of a child CT image and removing artifacts from the child CT image comprises: Decomposing the child CT image into different frequency sub-bands using wavelet transform; Detecting textures manifesting as high-frequency noise or anomalies in the decomposed subbands and thresholding the detected artifacts; The subbands after threshold processing are reconstructed back to the image space through inverse wavelet transform to obtain the CT image after artifact removal.

6. The method according to claim 1, characterized in that The registering of the artifact-removed pediatric CT plain scan image and the arterial phase image and the venous phase image comprises: The tumor segmentation result and the high-density bones in the abdomen are used to perform auxiliary feature point registration, wherein the registration evaluation standard adopts mutual information MI.

7. The method according to claim 6, characterized in that The auxiliary feature point registration using the tumor segmentation result and the high-density bones in the abdomen includes: The artifact-free pediatric CT plain scan images were normalized before registration using the Z score; The high-density bone is segmented by the Otsu adaptive threshold segmentation method, and the cavity is filled by a closing operation, and the edge points and corner points of the obtained tumor and high-density bone area are extracted respectively, wherein the edge points are detected by the Canny edge detection algorithm and the points are obtained by selecting points at equal intervals on the edge, and the corner points are obtained by the Harris corner point detection algorithm; Using a nearest neighbor matching algorithm to find corresponding feature point pairs in the artifact-removed pediatric CT plain scan image, the arterial phase image, and the venous phase image; Use affine transformation for registration and estimate the transformation parameters by minimizing the mean square error of matching point pairs; The similarity measure is defined using mutual information and feature point matching error; Maximize the similarity measure via gradient descent; Iteration is performed until the similarity measure no longer increases significantly and the feature point matching error is less than a preset value, then the iteration is stopped and the registration is completed.

8. The method according to claim 1, characterized in that The step of evaluating the enhancement degree by using the absolute enhancement value image and the normalized enhancement image comprises: Calculating enhancement distribution information of the patient's tumor area according to the normalized enhanced image; The proportion of the part of the tumor with enhancement greater than the set threshold to the total tumor volume was calculated.

9. A device for evaluating the activity of common solid tumors in the abdomen of children, characterized in that: include: Processor, memory; The memory is used to store computer programs; The processor is used to execute the method for active imaging assessment of common solid tumors in the abdomen of children as described in any one of claims 1 to 8 by calling the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for active imaging assessment of common solid tumors in the abdomen of children as described in any one of claims 1 to 8 is implemented.

Images