Prostate MRI graphic analysis method

Abstract

The invention relates to a prostate MRI graphic analysis method. The method comprises the following steps: S1, establishing a prostate malignant tumor clinical diagnosis medical case database; s2, a prostate MRI image full-automatic segmentation model is established and trained; s3, the input prostate MRI image is segmented; s4, shape features, first-order statistical features, texture features, wavelet features and depth features of the segmented prostate MRI image are extracted and screened; and S5, inputting the screened radiomics characteristics into a prostate malignant tumor analysis model to obtain an analysis result. The method has the advantages that through the E2C-Transform network, the prostate tumor MRI image can be smoothly processed, the characteristic area of the prostate can be enhanced and reserved, the image noise can be reduced to the maximum extent, and the accurate and high-quality segmentation of the prostate MRI image can be ensured; and an integrated algorithm combining the feature sparsity of the LASSO characteristics and the information screening advantage of the correlation entropy is adopted, so that the accuracy of the analysis model is enhanced, and the transparency and interpretability of the model decision are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital medical image processing and analysis, and particularly relates to a prostate MRI image analysis method. BACKGROUND

[0002] Prostate tumor has become a major health threat in the male urinary system, and its early diagnosis is particularly important. With the development of medical technology, medical imaging technology has gradually become the main means for the diagnosis of prostate tumors. MRI has the characteristics of no radiation, high resolution and high contrast for prostate tissue, and has gradually become the first choice for prostate cancer detection. However, the image features of prostate MRI are complex, and radiologists are required to undergo a large amount of training and process a large amount of multi-modal image data, which greatly increases their work pressure. Therefore, although MRI technology has brought great help to the diagnosis of prostate cancer, how to improve the accuracy and efficiency of prostate MRI image analysis and reduce the workload of doctors is still an important topic to be solved. SUMMARY

[0003] The present application mainly solves the above problems, and provides a prostate MRI image analysis method, which uses deep learning technology to realize accurate segmentation of prostate MRI images, and screens out the key core basis for diagnosis of prostate malignant tumors. Finally, a more accurate and individualized prostate malignant tumor analysis model is established by combining the ResNet network structure, which provides auxiliary decision-making for clinical practice.

[0004] The technical scheme adopted by the present application to solve the technical problems is a prostate MRI image analysis method, comprising the following steps:

[0005] S1: Establishing a prostate malignant tumor clinical diagnosis medical record database;

[0006] S2: Establishing a prostate MRI image full-automatic segmentation model and training;

[0007] S3: Segmenting the input prostate MRI image by using the prostate MRI image full-automatic segmentation model;

[0008] S4: Extracting and screening the shape features, first-order statistical features, texture features, wavelet features and deep features of the segmented prostate MRI image;

[0009] S5: Inputting the screened image features into the prostate malignant tumor analysis model to obtain an analysis result.

[0010] As a preferred scheme of the above scheme, the step S1 comprises: obtaining the MRI image of a historical prostate malignant tumor patient and manually outlining the contour of the prostate tumor region; and collecting the clinical, pathological and laboratory data of the patient corresponding to the MRI image.

[0011] As a preferred solution of the above-mentioned solution, the step S2 comprises the following steps:

[0012] S2.1: prostate multi-sequence MRI image registration, including registration of different sequence images of prostate MRI images; correcting the inhomogeneous field offset of the image; data standardization for each sequence image;

[0013] S2.2: establishing a prostate MRI image full-automatic segmentation model based on an E2C-Transformer network;

[0014] S2.3: training the prostate MRI image full-automatic segmentation model using the registered prostate MRI image.

[0015] As a preferred solution of the above-mentioned solution, the shape feature represents the shape of the prostate tumor in a triangular grid, and quantitatively evaluates the tumor shape, which includes flatness, elongation, volume, surface area, maximum diameter, surface volume ratio, density, eccentricity, sphericity, minimum axis length and second largest axis length.

[0016] As a preferred solution of the above-mentioned solution, the first-order statistical feature quantitatively describes the distribution of pixel gray level in the prostate tumor by statistical indicators: mean, median, interquartile range, gray value range, minimum value, maximum value, variance, root mean square, skewness, kurtosis, uniformity, and energy.

[0017] As a preferred solution of the above-mentioned solution, the texture feature is used to describe the heterogeneity of gray distribution in the prostate tumor, which is found by constructing various gray matrices, including 24 gray co-occurrence matrices, 16 gray run-length matrices, 14 gray dependence matrices, 16 gray region size matrices and 5 neighborhood gray tone difference matrices.

[0018] As a preferred solution of the above-mentioned solution, the wavelet feature is obtained by calculating the first-order statistical feature and the texture feature of the original image through eight decomposition methods, and different decomposition methods use different combinations of high-pass filtering or low-pass filtering to filter the x, y and z directions of the original image.

[0019] As a preferred solution of the above-mentioned solution, the deep feature is extracted by a CNN model, and the CNN model comprises a plurality of layers of alternating convolution layers and subsampling layers and at least one layer connection layer, and the i-th layer convolution feature map The calculation formula is:

[0020]

[0021] where l and h are the length and width of the convolution filter, respectively, and w jkis the learning weight of the convolution kernel, is the input of the convolution layer i, b i is the bias term of the i-th layer filter, f act (·) is an activation function;

[0022] The subsampling layer adopts a pooling layer to perform downsampling operation, and the calculation formula of the i-th pooling layer is:

[0023]

[0024] Wherein, κ(·) represents a pooling method, including a maximum pooling method, an average pooling method and a random pooling method, x i-1 is the input of the pooling layer i, x i is the output of the pooling layer i, is the bias term of the pooling layer i, f act (·) is an activation function;

[0025] The output vector of the i-th full connection layer is:

[0026] v i =f act (W i v i-1 +B i )

[0027] Wherein, v i-1 is the input of the i-th full connection layer, W i is the weight value of the network, B i is the bias term of the network, f act (·) is an activation function.

[0028] As a preferred scheme of the above scheme, the screening in the step S4 is performed by using an integrated algorithm based on LASSO and related entropy

[0029]

[0030] Wherein, the matrix X represents the imageomic features of a certain prostate malignant tumor patient, w is a weight vector, Y is the diagnosis result of the patient, and ρ and δ are experimental parameters,

[0031] ||2 is the norm of the vector L2, ||W|| 2,1 is the regularizer of the L2 norm on W.

[0032] As a preferred scheme of the above scheme, the prostate malignant tumor analysis model is established based on the network structure of ResNet18.

[0033] The advantages of the present application are: through E 2The C-Transformer network can smooth the prostate tumor MRI image, enhance and retain the feature area of the prostate, and minimize image noise, thereby ensuring accurate and high-quality segmentation of the prostate MRI image. The integrated algorithm combining the feature sparsity of LASSO and the information screening advantage of correlation entropy enhances the accuracy of the analysis model and ensures the transparency and interpretability of the model decision. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 FIG. 1 is a structural schematic diagram of an E2C-Transformer network.

[0035] Figure 2 FIG. 5 is a structural schematic diagram of a CTPC module.

[0036] Figure 3 FIG. 8 is a structural schematic diagram of a CNN model.

[0037] Figure 4 FIG. 12 is a structural schematic diagram of a ResNet18 network. DETAILED DESCRIPTION

[0038] The technical solutions of the present application will be further described below by means of examples and in conjunction with the accompanying drawings.

[0039] Embodiment:

[0040] The prostate MRI image analysis method of the present embodiment comprises the following steps:

[0041] S1: Establish a clinical diagnosis medical record database of prostate malignant tumors.

[0042] S1.1: Obtain the MRI image of a historical prostate malignant tumor patient and manually outline the prostate tumor area. In the present embodiment, the MRI image is a 1.5T and 3.0T MRI image using multiple MRI image sequences, including T2WI, DWI and ADC, and the outline of the prostate tumor area is obtained by manually outlining using a labeling software. In order to ensure the accuracy of the labeling result, the project team compares the outlining results of two members, and if there is inconsistency, the correct outline is confirmed through doctor consultation. In this way, subjective errors can be minimized and the quality of the labeled data can be improved.

[0043] S1.2: Collect clinical, pathological and laboratory data corresponding to the MRI image of the patient. In this embodiment, the clinical information includes height, weight, age, BSA, whether smoking, whether drinking and comorbidities (diabetes, hyperlipidemia, coronary heart disease, hypertension), survival time (months), survival status score and other clinical characteristics; pathological data includes Gleason score, immunohistochemistry; laboratory indicators include biochemistry (blood glucose, blood urea nitrogen, creatinine, uric acid, AST, ALT, LDH, alkaline phosphatase, ALP, albumin, etc.), testosterone, total PSA, free PSA, tumor stage (T stage, N stage, M stage).

[0044] S2: Establish a fully automatic prostate MRI image segmentation model and train it.

[0045] S2.1: Prostate multi-sequence MRI image registration, including registration of different sequence images of prostate MRI images; correcting the inhomogeneous field offset of the image; data standardization for each sequence image. In this embodiment, the FLIRT registration tool in the FSL software library is used to realize the registration of the T2WI, DWI and ADC three sequence images of the prostate MRI image, so as to ensure that the pixels at the same position correspond to the same prostate tissue. Secondly, the N4ITK method of ANTS software is used to correct the inhomogeneous field offset of the image, so as to improve the accuracy of the automatic segmentation algorithm. Finally, the Z-score method is used to standardize the data of each sequence image, so as to eliminate the contrast difference of different sequence images and enhance the contrast of the prostate tumor region and the non-tumor region.

[0046] S2.2: Establish a prostate MRI image fully automatic segmentation model based on E2C-Transformer network, as shown in Figure 1 The end-to-end CNN-Transformer structure is adopted, the network has a symmetric encoder-decoder structure, contains a skip connection, can fuse local and global features of different resolutions, and restores the resolution step by step through deconvolution. All convolutional layers in the network are provided with batch normalization layers and ReLU activation layers to avoid overfitting. In order to strengthen feature extraction, the network is designed with 8 consecutive convolutional layers at the bottom. In order to control the computational amount of the Transformer branch, the FFP selects different down-sampling steps at different resolutions, and the block embedding size of each level is kept at 4096, and the feature map size input to the Transformer is kept at 16x16x16. As shown in Figure 2As shown in the figure, the CTPC module of the E2C-Transformer network consists of three parts: a CNN branch, a Transformer branch, and an FFP. The CNN branch consists of two integrated convolutional modules, each of which includes a 1×1×1 down-projection convolutional layer, a 3×3×3 spatial convolutional layer, a 1×1×1 up-projection convolutional layer, and a residual connection. The CNN branch is used to extract continuous local features by sliding across the neighborhood graph, helping to preserve fine local features. Therefore, the CNN branch can continuously provide local details to the Transformer branch. The Transformer branch consists of a multi-head self-attention module and a multi-layer perceptron (MLP) block, with two layers of normalization and two residual connections. Due to computational constraints and feature map resolution limitations, the feature maps extracted by the CNN branch are downsampled to 16×16×16 block embeddings. Simultaneously, the FFP is used to account for shape differences, interactively coupling local and global features. Through the interaction between the CNN and Transformer branches, the E2C-Transformer network can better capture the local and global features of prostate tumors, alleviating the problem of high diversity in tumor shape, structure, and location. To address the class imbalance in prostate tumor MRI images, this project extends the generalized Dice loss function to the prostate tumor region. Weights are assigned based on the volume ratio of the prostate tumor region to the background, giving higher weights to regions with smaller classes to reduce the model's focus on background regions. Furthermore, this project employs TTA and volume-constrained post-processing to achieve more accurate prostate MRI image segmentation results.

[0047] S2.3: Use the registered prostate MRI images to train the fully automatic prostate MRI image segmentation model.

[0048] S3: Segment the input prostate MRI image using the fully automatic prostate MRI image segmentation model;

[0049] S4: shape features, first-order statistics features, texture features, wavelet features and deep features of the segmented prostate MRI images are extracted and screened. The shape features mainly describe the shape of the target region and its related geometric properties. The shape features represent the prostate tumor shape with a triangular mesh, quantitatively evaluate the tumor shape, and the features include flatness, elongation, volume, surface area, maximum diameter, surface-to-volume ratio, density, eccentricity, sphericity, minimum axis length and second largest axis length. The first-order statistics features focus on the distribution of intensity values of individual voxels in the image, and do not involve their positions or relationships in space. The first-order statistics features quantitatively describe the distribution of pixel gray level in the prostate tumor by statistical indicators: mean, median, interquartile range, gray value range, minimum value, maximum value, variance, root mean square, skewness, kurtosis, uniformity, and energy. The texture features are used to describe the heterogeneity of the gray distribution in the prostate tumor. Various gray matrices are constructed to discover the heterogeneity, including 24 gray co-occurrence matrices, 16 gray run-length matrices, 14 gray dependence matrices, 16 gray region size matrices and 5 neighborhood gray tone difference matrices. The wavelet features are obtained by calculating the first-order statistics features and texture features of the original image through 8 decomposition methods. Different decomposition methods use different combinations of high-pass filtering or low-pass filtering to filter the x, y and z directions of the original image. Let L and H be the low-pass filtering and high-pass filtering, respectively. The wavelet decomposition of the image can be marked as LLL, LLH, LHL, LHH, HLL, HLH, HHL and HHH. For example, HHL is obtained by high-pass filtering in the x direction, high-pass filtering in the y direction and low-pass filtering in the z direction of the original image. The decomposition image obtained has the same size as the original image. All first-order statistics features and all texture features are calculated for the 8 decomposition methods to obtain the corresponding wavelet features.

[0050] The deep features are extracted by a CNN model, which includes a plurality of layers of alternating convolutional layers and subsampling layers and at least one layer connection layer. In this embodiment, the CNN model is as shown in FIG. 2, which includes two convolutional layers, two subsampling layers and one fully connected layer. Figure 3

[0051] The i-th layer of convolutional feature map The calculation formula is:

[0052]

[0053] wherein, l and h are the length and width of the convolutional filter, w jk is the learning weight of the convolutional kernel, is the input of the convolutional layer i, b i is the bias term of the i-th layer filter, f act (·) is an activation function.

[0054] ​The subsampling layer uses the pooling layer for downsampling operation. The calculation formula of the i-th pooling layer is:

[0055]

[0056] Among them, κ(·) represents the pooling method, including maximum pooling, average pooling and random pooling, x i-1 is the input of pooling layer i, x i is the output of pooling layer i, is the bias term of pooling layer i, f act (·) is the activation function;

[0057] The output vector of the i-th fully connected layer is:

[0058] v i =f act (W i v i-1 +B i )

[0059] Among them, v i-1 is the input of the i-th fully connected layer, W i is the weight value of the network, B i is the bias term of the network, f act (·) is the activation function.

[0060] In addition, in step S4, an integrated algorithm based on LASSO and correlation entropy is used for screening

[0061]

[0062] Among them, the matrix X represents the imaging features of a patient with prostate malignancy, w∈W is the weight vector, Y is the patient's diagnosis result, ρ and δ are experimental parameters, ||||2 is the norm of the vector L2, ||W|| 2,1 is the L2 norm regularizer on W. Regularizer ||W|| 2,1 By penalizing W with the coefficient of incentive row sparsity to select the corresponding features in X, the correlation between features with stronger recognition ability can be learned by inputting features with larger weights into the prostate malignancy analysis model, thereby improving the accuracy of the analysis.

[0063] S5: Input the selected imaging features into the prostate malignancy analysis model to obtain analysis results. The prostate malignancy analysis model is built based on the ResNet18 network structure. Figure 4As shown, the network structure of ResNetl 8 starts with a 7x7 convolutional layer with 64 filters and a stride of 2, which aims to capture the basic features of the image. This is followed by a 3x3 max-pooling layer with a stride of 2, which reduces the spatial dimension of the features and improves the computational efficiency. The main body of the network includes four blocks, each of which has two residual units. The number of filters in these residual units gradually increases from 64 to 128, 256, and finally to 512, enabling the network to capture features at different levels of abstraction. After these blocks, a global average pooling layer is used to reduce the number of parameters and provide input to the fully connected layer, which outputs the required number of classes.

[0064] The specific embodiments described herein are merely illustrative of the spirit of the application. Various modifications or changes in the specific embodiments described herein can be made by those skilled in the art without departing from the spirit of the application, and the application is not limited to the specific embodiments described herein.

Claims

1. A method of prostate MRI pattern analysis, characterized by: The method comprises the following steps: S1: establishing a clinical diagnosis medical record database of prostate malignant tumors; S2: establishing a prostate MRI image full-automatic segmentation model and training; S3: segmenting the input prostate MRI image by using the prostate MRI image full-automatic segmentation model; S4: extracting shape features, first-order statistical features, texture features, wavelet features and deep features of the segmented prostate MRI image and screening; S5: inputting the screened radiomics features into a prostate malignant tumor analysis model to obtain an analysis result.

2. The method of prostate MRI pattern analysis of claim 1, wherein: The step S1 comprises: obtaining MRI images of historical prostate malignant tumor patients and manually outlining the contours of the prostate tumor regions; and collecting clinical, pathological and laboratory data of the patients corresponding to the MRI images.

3. The method of claim 1, wherein: The step S2 comprises the following steps: S2.1: prostate multi-sequence MRI image registration, comprising: registering different sequence images of the prostate MRI image; correcting the inhomogeneous field offset of the image; and performing data standardization on each sequence image; S2.2: establishing a prostate MRI image full-automatic segmentation model based on an E2C-Transformer network; S2.3: training the prostate MRI image full-automatic segmentation model by using the registered prostate MRI image.

4. The method of claim 1, wherein: The shape features represent the shape of the prostate tumor in a triangular mesh, and quantitatively evaluate the shape of the tumor, including flatness, elongation, volume, surface area, maximum diameter, surface-to-volume ratio, density, eccentricity, sphericity, minimum axis length and second largest axis length.

5. The method of claim 1, wherein: The first-order statistical features quantitatively describe the distribution of pixel gray level in the prostate tumor by statistical indicators: mean, median, interquartile range, gray value range, minimum value, maximum value, variance, root mean square, skewness, kurtosis, uniformity and energy.

6. The method of claim 1, wherein: The texture features are used to describe the heterogeneity of the gray distribution in the prostate tumor, which are found by constructing various gray matrices, including 24 gray level co-occurrence matrices, 16 gray run length matrices, 14 gray dependence matrices, 16 gray region size matrices and 5 neighborhood gray tone difference matrices.

7. The method of claim 1, wherein: The wavelet features are obtained by calculating the first-order statistical features and texture features of the original image through eight decomposition methods, and different decomposition methods adopt different combinations of high-pass filtering or low-pass filtering on the x, y and z directions of the original image.

8. The method of claim 1 wherein: The deep features are extracted by a CNN model, the CNN model comprises a plurality of layers of alternating convolution layers and sub-sampling layers and at least one layer connection layer, the i-th layer convolution feature map The calculation formula is: where l and h are the length and width of the convolutional filter, respectively, w jk is the learning weight of the convolution kernel, is the input of the convolutional layer i, b i is the bias term of the i-th layer filter, f act (·) is the activation function; The sub-sampling layer adopts a pooling layer to perform down-sampling operation, and the calculation formula of the i-th pooling layer is: where κ(·) represents the pooling method, including the max pooling method, the average pooling method and the random pooling method, x i-1 is the input of the pooling layer i, x i is the output of the pooling layer i, is the bias term of the pooling layer i, f act (·) is the activation function; the output vector of the i-th layer full connection layer is: v i = f act (W i v i-1 + B i ) wherein v i-1 is the input of the i-th full connection layer, W i is the weight value of the network, B i is the bias term of the network, and f act is the activation function.

9. The method of claim 1 wherein: The screening in the step S4 is performed by using an integrated algorithm based on LASSO and correlation entropy wherein the matrix X represents the imaging omics features of a certain prostate cancer patient, w e W is a weight vector, Y is the diagnosis result of the patient, p, d are experimental parameters, |||2 is the norm of vector L2, ||W||2 is the norm of vector W 2,1 is the regularizer of L2 norm on W.

10. The method of claim 1 wherein: The prostate malignant tumor analysis model is established based on the network structure of ResNet18.

Images