Method and device for automatically identifying benign and malignant kidney cystic lesions based on magnetic resonance image

Abstract

The invention relates to a method and a device for automatically identifying benign and malignant kidney cystic lesions based on magnetic resonance images. The method comprises the following steps: S1, preprocessing and standardizing a T2 weighted image, a diffusion weighted image, an apparent diffusion coefficient image, a T1 weighted image, a skin medullary phase image, a parenchyma phase image and an excretion phase image; s2, respectively training automatic segmentation models corresponding to different images in a targeted manner, and predicting a focus by using the automatic segmentation models; s3, extracting morphological features, first-order features and textural features of the lesions from all the lesions, wherein the morphological features, the first-order features and the textural features comprise features of capsule walls, partitions and nodules of the lesions; screening the extracted features, and constructing a classification model by using the features with good robustness; and S4, preprocessing and standardizing the image of the current patient, respectively inputting the image into each corresponding automatic segmentation model, and operating the segmentation model and the classification model to realize benign and malignant recognition based on image recognition. According to the invention, integrated and automatic benign and malignant accurate diagnosis of kidney cystic lesions is realized.

Description

Technical Field

[0001] This invention relates to magnetic resonance renal image processing technology in the medical field, and particularly to a method and apparatus for automatic identification of benign and malignant renal cystic lesions (referring to renal lesions that are predominantly fluid-based in images, with less than 25% solid component enhancement) based on magnetic resonance images. Background Technology

[0002] Cystic renal masses (CRMs) are extremely common in clinical practice, with a prevalence as high as 50% in people over 50 years of age. CRMs encompass various pathological types, and the treatment plans and survival prognoses differ depending on the subtype. Diagnosing simple renal cysts is relatively easy, but when lesions present with complex features such as wall thickening, septa, irregular enhancement, or nodules, accurate preoperative assessment of benign or malignant characteristics becomes extremely challenging. While the widely used Bosniak classification system helps assess malignancy risk, its classification criteria are cumbersome, dependent on physician experience, and suffer from significant inter- and intra-observer inconsistencies, limiting its standardized clinical application. In recent years, artificial intelligence has shown great potential in the classification and malignancy risk assessment of CRMs. However, due to technical difficulties and data challenges, there are currently no reports of automated segmentation for CRMs. Most existing methods still rely on manual delineation for lesion segmentation and information extraction, which is not only time-consuming and labor-intensive but also prone to introducing subjective bias. Summary of the Invention

[0003] To address the problems in the prior art, this application proposes an automatic method for identifying benign and malignant renal cystic lesions using magnetic resonance imaging, comprising: S1: Preprocess and standardize the training images. The preprocessing and standardization include resampling and data augmentation of the training images. The training images include T2-weighted images, diffusion-weighted images, apparent diffusion coefficient images, T1-weighted images, corticomedullary phase images, parenchymal phase images, and excretion phase images. S2: Based on the 3D V-Net convolutional neural network, automatic segmentation models are trained for T2-weighted images, diffusion-weighted images, apparent diffusion coefficient images, T1-weighted images, corticomedullary phase images, parenchymal phase images, and excretion phase images, and each automatic segmentation model is used to predict lesions. S3: Extract morphological features, first-order features and texture features of lesions from the lesions predicted by each automatic segmentation model, including features of lesion cyst walls, septa and nodules; screen the extracted features to retain the features most relevant to the differentiation of benign and malignant lesions, and construct a classification model of benign and malignant renal cystic lesions based on random forest using the features most relevant to the differentiation of benign and malignant lesions; the screening of the extracted features includes the following three-stage feature screening: (1) Perform variance thresholding: select a threshold of 0.005 to remove features with low variability; (2) Perform mutual information method: focus on the 100 features most relevant to the pathological label; (3) Perform minimum absolute contraction and selection operator regression: use 10-fold cross-validation and 50 logarithmic scale α values ​​to retain features with absolute coefficient values ​​greater than 0.01, and compress the coefficients of irrelevant features to zero through L1 regularization; S4: Preprocess and standardize the current patient's T2-weighted image, diffusion-weighted image, apparent diffusion coefficient image, T1-weighted image, corticomedullary phase image, parenchymal phase image, and excretion phase image. Then, input the preprocessed and standardized images into the corresponding automatic segmentation models and run the automatic segmentation models and the random forest-based benign and malignant classification model for renal cystic lesions to achieve benign and malignant identification of renal cystic lesions based on image recognition.

[0004] According to some embodiments of the present invention, the data enhancement includes pixel intensity normalization; the pixel intensity normalization includes: based on the signal intensity characteristics of the lesion on each magnetic resonance sequence, a relatively effective lesion signal intensity range is preset for different magnetic resonance sequences, the signal intensity values ​​of this range are mapped to between 0 and 1, and the intensity values ​​outside the range are clipped.

[0005] According to some embodiments of the present invention, the 3D V-Net convolutional neural network includes an encoder and a decoder; the encoder uses a 3D convolutional layer with a 2×2×2 convolutional kernel and a stride of 2, and each downsampling layer also extracts hierarchical features through multiple convolution operations; the decoder gradually restores the spatial resolution by using a 3D transposed convolutional operation with a 2×2×2 convolutional kernel and a stride of 2, and splices it with the features of the corresponding layer of the encoder through skip connections; the final output layer uses a 5×5×5 convolutional kernel to map the 32-channel features into a 2-channel output, representing the lesion and the background respectively.

[0006] According to some embodiments of the present invention, the parameters of the random forest-based classification model for benign and malignant renal cystic lesions are set as follows: class weights (0:0.3, 1:0.75), Gini coefficient as splitting criterion, maximum depth 12, feature selection method sqrt, minimum number of split samples 10, and number of weak learners 20.

[0007] According to some embodiments of the present invention, the automatic segmentation network corresponding to the automatic segmentation model uses the parameter-corrected linear unit (PReLu) as the activation function.

[0008] According to some embodiments of the present invention, the method further includes introducing a Dropout layer in the segmentation model for regularization to prevent overfitting.

[0009] According to some embodiments of the present invention, the resampling includes: calculating the ratio between the old and new voxel spacing, generating a new coordinate grid based on the scaling ratio of each dimension, and performing bilinear interpolation on the original data, and estimating the intermediate value by linearly combining the four nearest neighbor data points.

[0010] According to some embodiments of the present invention, the morphological features, first-order features and texture features of lesions extracted from all lesions predicted by various automatic segmentation models are preprocessed. The preprocessing includes standardizing the morphological features, first-order features and texture features using StandardScaler, with missing values ​​filled with zeros.

[0011] This application also proposes an automatic identification device for benign and malignant renal cystic lesions in magnetic resonance imaging, which includes a processor and a memory; the processor executes the method for automatic identification of benign and malignant renal cystic lesions in magnetic resonance imaging based on a computer program stored in the memory.

[0012] This application also proposes a computer program product, including a computer program, wherein when the computer program is executed by a processor, it implements the steps of an automatic identification method for benign and malignant renal cystic lesions in magnetic resonance images.

[0013] The above embodiments of the present invention propose for the first time to combine automatic segmentation technology of renal cystic lesions in magnetic resonance images with machine learning classification algorithms for renal cystic lesions in magnetic resonance images, establish a multidimensional feature extraction and three-stage progressive feature screening system that integrates morphology, statistics and texture, and integrate it on the random forest learning algorithm, so as to realize the integrated, automated and accurate diagnosis of benign and malignant renal cystic lesions in magnetic resonance images, effectively overcome the shortcomings of traditional methods in terms of efficiency and consistency, and have great clinical application value. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used are briefly described below: Figure 1 A flowchart of a method for automatic identification of benign and malignant renal cystic lesions from magnetic resonance images according to some embodiments of the present invention; Figure 2 This is a schematic diagram of an automatic segmentation 3D V-Net convolutional neural network architecture according to some embodiments of the present invention; Figure 3 This is a schematic diagram of a machine learning algorithm for classifying benign and malignant diseases according to some embodiments of the present invention; Figure 4 This is a schematic diagram of a model integration strategy according to some embodiments of the present invention; Figure 5 This is an example of the automatic segmentation and benign / malignant diagnosis results of the automatic identification method for benign / malignant renal cystic lesions of the present invention on multiple MR sequences; Figure 6 This is a schematic structural diagram of an automatic identification device for benign and malignant renal cystic lesions in magnetic resonance images according to some embodiments of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0016] Before introducing the technical solution, let's first explain some of the terms used in this application: Convolutional Neural Networks (CNNs): A type of feedforward neural network that includes convolutional computations and has a deep structure. They are widely used in image recognition and vision tasks and are one of the core models in deep learning.

[0017] Automatic segmentation model: Based on a convolutional neural network, this mathematical model is trained on labeled samples and can automatically generate segmentation results for target regions in an input image.

[0018] Random Forest: A supervised learning algorithm based on the Bagging ensemble strategy, mainly used for classification and regression tasks. It completes the task by constructing and combining multiple decision trees, aiming to achieve better generalization performance, robustness and accuracy than a single model.

[0019] Machine learning: a technique that uses algorithms and models to enable computers to automatically learn from data and make predictions or decisions; it is a branch of artificial intelligence.

[0020] Figure 1 The overall flow of the automatic identification method for renal cystic lesions in this invention is shown, and the method includes the following steps S1-S4: S1: Preprocess and standardize the training images. The preprocessing and standardization include resampling and data augmentation of the training images. The training images include T2-weighted images (T2WI images), diffusion-weighted images (DWI images), apparent diffusion coefficient images (ADC images), T1-weighted images (T1WI images), corticomedullary phase images (CP images), parenchymal phase images (NP images), and excretory phase images (EP images).

[0021] Unlike existing technologies, in this invention, T2WI images, DWI images, ADC images, T1WI images, CP images, NP images, and EP images are selected as training images for the corresponding automatic segmentation models according to the purpose of the invention, so as to give full play to the imaging advantages of different magnetic resonance sequences and improve the accuracy of recognition results.

[0022] Furthermore, in response to the problem that the quality of MR images is easily affected by equipment and scanning schemes in existing technologies, this invention employs resampling and data augmentation methods to preprocess and standardize images, so as to eliminate image inconsistencies, enhance data diversity, and improve the model's generalization ability.

[0023] According to some embodiments of the present invention, the voxel spacing of all images is uniformly resampled to 1.0 mm × 1.0 mm × 1.0 mm. The resampling process includes: calculating the ratio between the old and new voxel spacings; generating a new coordinate grid based on the scaling ratio of each dimension; and interpolating the original data using bilinear interpolation. The intermediate value is estimated by linearly combining the four nearest neighbor data points, thereby completing data resampling while maintaining local features and spatial coherence. Image boundaries are padded or truncated to avoid interpolation overflow.

[0024] According to some embodiments of the present invention, the data augmentation includes expanding the training samples and greatly improving the robustness of the model through affine transformation, noise reduction, pixel intensity normalization, random cropping, rotation and flipping.

[0025] According to some embodiments of the present invention, the pixel intensity normalization includes: based on the signal intensity characteristics of the lesion on each magnetic resonance sequence, pre-setting a relatively effective lesion signal intensity range for each magnetic resonance sequence, mapping the signal intensity values ​​within this range to between 0 and 1, and cropping the intensity values ​​outside the range. The specific pre-set lesion signal intensity ranges for different magnetic resonance sequences are as follows: ADC image: 800-3600; CP image: 50-600; DWI image: 150-650; EP image: 50-1000; NP image: 100-1500; T1WI image: 50-600; T2WI image: 150-1800. In traditional pixel intensity normalization processing during data preprocessing, a global maximum-minimum normalization method is used, but this method is susceptible to interference from extreme noise values, compressing the contrast of key tissues. The present invention, by adopting the aforementioned pixel intensity normalization processing method, fully considers the specific signal intensity range of lesions in different sequences. It not only effectively suppresses the interference of background noise and artifacts, but also clearly highlights the gray level difference between the target lesion and the surrounding normal tissue. This provides high-quality, high signal-to-noise ratio input data for subsequent targeted training of segmentation models for each sequence, ensuring that the lesion features under each sequence are fully captured and utilized.

[0026] S2: Based on the 3D V-Net convolutional neural network, automatic segmentation models are trained for T2WI, DWI, ADC, T1WI, CP, NP and EP images respectively, and the lesions are predicted using each automatic segmentation model.

[0027] The 3D V-Net convolutional neural network here is specifically designed for 3D medical imaging, and its network structure is as follows: Figure 2 As shown, the network consists of an encoder (downsampling path) and a decoder (upsampling path). The decoder concatenates features from the corresponding layer of the encoder through skip connections to achieve multi-scale feature fusion, thereby promoting detail recovery and integration of contextual information.

[0028] According to some embodiments of the present invention, the encoder employs a 3D convolutional layer with a 2×2×2 convolutional kernel and a stride of 2, which halves the spatial resolution of the feature map while doubling the channel depth. Each downsampling layer also extracts hierarchical features through multiple convolution operations. The decoder gradually restores the spatial resolution by using a 3D transposed convolutional operation with a 2×2×2 convolutional kernel and a stride of 2, and concatenates the features with the corresponding layers of the encoder through skip connections, thereby reconstructing detailed information.

[0029] The automatic segmentation model of this invention uses the Parametric Corrected Linear Unit (PReLU) as the activation function. Compared with the traditional Parametric Corrected Linear Unit (ReLU) activation function, it has adjustable parameters and can have better gradient propagation characteristics and negative value handling capabilities.

[0030] The present invention may further include introducing a Dropout layer into the segmentation model for regularization to effectively prevent overfitting.

[0031] The final output layer of this invention uses a 5×5×5 convolution kernel to map 32-channel features into 2-channel outputs, representing the foreground (lesion) and background respectively.

[0032] S3: Extract morphological features, first-order features and texture features of lesions from the lesions predicted by each automatic segmentation model, including features of lesion cyst walls, septa and nodules; screen the extracted features to retain the features most relevant to the differentiation of benign and malignant lesions, and construct a classification model of benign and malignant renal cystic lesions based on random forest using the features most relevant to the differentiation of benign and malignant lesions; the screening of the extracted features includes the following three-stage feature screening: (1) Perform variance thresholding: select a threshold of 0.005 to remove features with low variability; (2) Perform mutual information method: focus on the 100 features most relevant to the pathological label; (3) Perform minimum absolute contraction and selection operator regression: use 10-fold cross-validation and 50 logarithmic scale α values ​​to retain features with absolute coefficient values ​​greater than 0.01, and compress the coefficients of irrelevant features to zero through L1 regularization.

[0033] In step S3, constructing a classification model for benign and malignant renal cystic lesions based on random forest (hereinafter referred to as the classification model) may include: performing inter-group difference analysis on all selected features, and finally constructing a classification model using the random forest algorithm based on features showing significant differences between the two groups. The inter-group difference analysis includes first clearly classifying all lesions into benign or malignant groups according to the "gold standard" (surgical pathological diagnosis), and then examining each feature one by one to determine whether the value of that feature is statistically significantly different between the two groups. (Reference) Figure 3 This diagram illustrates a learning algorithm for a random forest-based classification model of benign or malignant renal cystic lesions according to some embodiments of the present invention. The algorithm comprises multiple decision trees, each trained independently to predict the lesion's classification (benign or malignant). Through a compensatory effect of aggregating the predictions from the entire set of decision trees, the classification model effectively reduces the probability of incorrect predictions from individual decision trees. According to some embodiments of the present invention, setting the number of random seeds to 42 ensures the reproducibility of the experiment.

[0034] The operation process of the classification model mainly includes: three-stage feature selection, model training, and performance evaluation of the extracted morphological features, first-order features, and texture features. Figure 4 As shown, the three-stage feature screening includes (1) performing the variance threshold method: selecting a threshold of 0.005 to remove features with low variability (close to constant), which can effectively reduce the computational burden; (2) performing the mutual information method: focusing on the 100 features most strongly correlated with the pathological label to achieve significant dimensionality reduction; (3) performing minimum absolute contraction and selection operator regression: using 10-fold cross-validation and 50 log-scale α values ​​(10⁻ 4 Up to 10 0Features with absolute coefficients greater than 0.01 are retained, and L1 regularization is used to compress irrelevant feature coefficients to zero. This addresses feature collinearity while balancing model performance and generalization ability, ultimately yielding the optimal and sparsest feature subset. Although the features selected in the first two steps are all related to pathological labels, there may be information redundancy or collinearity among them. For example, the information provided by "maximum lesion diameter" and "lesion volume" is redundant. In this case, the system automatically retains one feature during minimum absolute shrinkage and selection operator regression screening, eliminating redundancy and obtaining a more concise and independent feature set. Through this three-stage feature screening, robust and interpretable features can be selected, ensuring computational feasibility while balancing linear and nonlinear relationships. The final feature set not only has strong predictive performance but is also concise, facilitating subsequent biological interpretation or clinical validation by researchers. Based on the imbalanced distribution of benign and malignant cases in clinical practice, according to some embodiments of the present invention, the classification model parameters are set as follows: class weights (0:0.3, 1:0.75), Gini coefficient as the splitting criterion, maximum depth 12, feature selection method sqrt, minimum number of split samples 10, and number of weak learners 20. This scheme can reduce the risk of overfitting while ensuring the retention of the most discriminative features. The performance evaluation of the model mainly includes the evaluation of diagnostic efficacy, which can be achieved using receiver operating characteristic (ROC) curve analysis.

[0035] According to some embodiments of the present invention, the method further includes preprocessing the morphological features, first-order features and texture features extracted from all lesions predicted by various automatic segmentation models. The preprocessing includes standardizing the morphological features, first-order features and texture features using StandardScaler, with missing values ​​filled with zeros.

[0036] This application also proposes using MRI data from external medical centers as an independent external test set to more objectively evaluate the diagnostic performance of the model.

[0037] According to some embodiments of the present invention, MRI data from three external medical centers are used as independent validation sets to comprehensively evaluate the model's performance on segmentation and classification tasks. The external validation set data is completely independent of the training set data, helping to avoid any form of data leakage and ensuring the objectivity and credibility of the validation results. Data augmentation is strictly prohibited during the validation process, and other processes in the preprocessing are completely consistent with the training phase. Segmentation performance is quantitatively evaluated using the Dice similarity coefficient (DSC). The closer the value is to 1, the closer the segmentation result is to the true situation, and the higher the segmentation accuracy. The benign / malignant classification performance is comprehensively analyzed using the area under the ROC curve (AUC), sensitivity, specificity, and confusion matrix.

[0038] Figure 5 The automatic segmentation and benign / malignant identification results of the renal cystic lesion automatic identification method of the present invention on multiple magnetic resonance imaging (MRI) sequences are shown, where red and green outlines represent the lesion boundaries manually labeled and predicted by the model, respectively. It can be seen that this method efficiently and accurately diagnoses renal cystic lesions.

[0039] S4: Preprocess and standardize the current patient's T2WI, DWI, ADC, T1WI, CP, NP, and EP images, and input the preprocessed and standardized images into the corresponding automatic segmentation models. Run each automatic segmentation model and the random forest-based benign and malignant classification model for renal cystic lesions to achieve benign and malignant identification of renal cystic lesions based on image recognition.

[0040] The current patient is one who needs to identify and diagnose the benign or malignant nature of renal cystic lesions. After the aforementioned steps S1-S3, the present invention can obtain a complete diagnostic model for automatic segmentation and classification of renal cystic lesions. Therefore, in this step, the current patient's T2WI, DWI, ADC, T1WI, CP, NP, and EP images are preprocessed and standardized, and the preprocessed and standardized images are input into the corresponding automatic segmentation models. By running each automatic segmentation model and the random forest-based classification diagnostic model, the benign or malignant diagnosis of renal cystic lesions based on automatic image recognition can be achieved, that is, an integrated automatic segmentation and automatic classification diagnosis of renal cystic lesions can be realized.

[0041] like Figure 5 As shown, after prediction by the automatic segmentation model, feature extraction, and classification model, the final predicted value is 0.349. Since it is less than 0.632 (this value is determined by the maximum Youden index of the integrated model's predictive effectiveness), the model is judged to be benign.

[0042] Figure 6 A schematic structural diagram of an automatic identification device 500 for benign and malignant renal cystic lesions using magnetic resonance imaging, illustrating some embodiments of the present invention. Figure 6 As shown, the device includes a processor 51, a memory 52, and a bus 53.

[0043] In some instances, the device may further include an input device 501, an input port 502, an output port 503, and an output device 504. The input port 502, processor 51, memory 52, and output port 503 are interconnected via a bus 53. The input device 501 and output device 504 are connected to the bus 53 via the input port 502 and output port 503, respectively, and thus connected to other components of the device. It should be noted that the output and input interfaces here can also be represented by I / O interfaces. Specifically, input device 501 receives input information from external sources, such as T2WI, DWI, ADC, T1WI, CP, NP, and EP images of renal cystic lesions, and transmits the input information to processor 51 through input port 502. Processor 51 processes the input information based on computer-executable instructions stored in memory 52 to generate output information, temporarily or permanently stores the output information in memory 52, and then transmits the output information to output device 504 through output port 503. Output device 504 outputs the output information to the outside of the device. The output information may be, for example, a benign or malignant result.

[0044] The aforementioned memory 52 includes a large-capacity memory for data or instructions.

[0045] Bus 53 includes hardware, software, or both, which couples multiple components together.

[0046] The processor 51 executes an automatic identification method for benign and malignant renal cystic lesions based on the computer program stored in the memory 52.

[0047] This application also proposes a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the method described herein.

[0048] The detailed descriptions of the figures mentioned above are included here by reference and will not be repeated here.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art will recognize that the present invention can have other modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. An automatic method for identifying benign and malignant renal cystic lesions from magnetic resonance imaging, characterized in that, include: S1: Preprocess and standardize the training images. The preprocessing and standardization include resampling and data augmentation of the training images. The training images include T2-weighted images, diffusion-weighted images, apparent diffusion coefficient images, T1-weighted images, corticomedullary phase images, parenchymal phase images, and excretion phase images. S2: Based on the 3D V-Net convolutional neural network, automatic segmentation models are trained for T2-weighted images, diffusion-weighted images, apparent diffusion coefficient images, T1-weighted images, corticomedullary phase images, parenchymal phase images, and excretion phase images, and each automatic segmentation model is used to predict lesions. S3: Extract morphological features, first-order features, and texture features of lesions from the lesions predicted by each automatic segmentation model, including features of lesion cyst walls, septa, and nodules; The extracted features are screened to retain the features most relevant to the identification of benign and malignant lesions. The features most relevant to the identification of benign and malignant lesions are used to construct a classification model of benign and malignant renal cystic lesions based on random forest. The screening of the extracted features includes the following three-stage feature screening: (1) Performing the variance threshold method: Select a threshold of 0.005 to remove features with low variability. (2) Perform the mutual information method: focus on the 100 features most strongly correlated with pathological labels; (3) Perform minimum absolute shrinkage and selection operator regression: use 10-fold cross-validation and 50 log scale α values, retain features with absolute values ​​of coefficients greater than 0.01, and compress unrelated feature coefficients to zero through L1 regularization; S4: Preprocess and standardize the current patient's T2-weighted image, diffusion-weighted image, apparent diffusion coefficient image, T1-weighted image, corticomedullary phase image, parenchymal phase image, and excretion phase image. Then, input the preprocessed and standardized images into the corresponding automatic segmentation models and run the automatic segmentation models and the random forest-based benign and malignant classification model for renal cystic lesions to achieve benign and malignant identification of renal cystic lesions based on image recognition.

2. The method for automatic identification of benign and malignant renal cystic lesions according to claim 1, characterized in that, The data enhancement includes pixel intensity normalization; The pixel intensity normalization includes: based on the signal intensity characteristics of the lesion on each magnetic resonance sequence, a relatively effective lesion signal intensity range is preset for different magnetic resonance sequences, the signal intensity values ​​of this range are mapped to between 0 and 1, and the intensity values ​​outside the range are clipped.

3. The method for automatic identification of benign and malignant renal cystic lesions according to claim 1, characterized in that, The 3D V-Net convolutional neural network includes an encoder and a decoder; the encoder uses a 2×2×2 convolutional kernel and a stride of 2 in a 3D convolutional layer, and each downsampling layer also extracts hierarchical features through multiple convolution operations; The decoder gradually restores the spatial resolution by using a 2×2×2 convolution kernel and a 3D transposed convolution operation with a stride of 2, and splices the features of the corresponding layer of the encoder through skip connections; the final output layer uses a 5×5×5 convolution kernel to map the 32-channel features into 2-channel outputs, representing lesions and background respectively.

4. The method for automatic identification of benign and malignant renal cystic lesions according to claim 3, characterized in that, The parameters of the random forest-based classification model for benign and malignant renal cystic lesions are set as follows: class weights (0:0.3, 1:0.75), Gini coefficient as the splitting criterion, maximum depth 12, feature selection method sqrt, minimum number of split samples 10, and number of weak learners 20.

5. The method for automatic identification of benign and malignant renal cystic lesions according to claim 1, characterized in that, The automatic segmentation network corresponding to the automatic segmentation model uses the parameter-corrected linear unit (PReLu) as the activation function.

6. The method for automatic identification of benign and malignant renal cystic lesions according to claim 1, characterized in that, It also includes introducing a Dropout layer in the segmentation model for regularization to prevent overfitting.

7. The method for automatic identification of benign and malignant renal cystic lesions according to claim 1, characterized in that, The resampling process includes: calculating the ratio between the old and new voxels, generating a new coordinate grid based on the scaling ratio of each dimension, and using bilinear interpolation to interpolate the original data, estimating the intermediate value by linearly combining the four nearest neighbor data points.

8. The method for automatic identification of benign and malignant renal cystic lesions according to claim 1, characterized in that, The morphological features, first-order features, and texture features of the lesions extracted from all lesions predicted by various automatic segmentation models are preprocessed. The preprocessing includes standardizing the morphological features, first-order features, and texture features using StandardScaler, with missing values ​​filled with zeros.

9. An automatic identification device for benign and malignant renal cystic lesions from magnetic resonance imaging, comprising a processor and a memory; the processor executing the method according to any one of claims 1-8 based on a computer program stored in the memory.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-8.

Images