Method for classifying and identifying lesions in magnetic resonance image based on causal line network model
By extracting and optimizing the lesion features of magnetic resonance imaging through a causal circuit network model, the problem of incomplete prediction of lesion signs in existing technologies is solved, and highly accurate and explainable lesion identification is achieved to assist clinical diagnosis.
Patent Information
- Application Number
- CN202510474927.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-16
- Publication Date
- 2025-09-16
AI Technical Summary
Existing research on automatic diagnosis of lesions based on magnetic resonance imaging is often based on only partial scanning sequences, making it difficult to comprehensively and accurately predict multiple lesion signs. In addition, general deep learning models lack interpretability and cannot effectively process causal correlation signals of multimodal input and output.
The design is based on a causal circuit network model, including an image encoder, a causal circuit module, a signal encoder, a self-attention module, and a causal relay module. Through training on multi-sequence magnetic resonance imaging datasets, it extracts lesion features, decouples and fuses information, and optimizes the prediction of lesion types and sign signals.
It improves the accuracy and interpretability of lesion type identification, can comprehensively predict lesion sign signals, provide a reliable basis for clinical diagnosis, and enhance the interpretability and prediction accuracy of the model.
Smart Images
Figure CN120655957A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a method for classifying and identifying lesions in magnetic resonance images based on a causal circuit network model. Background Art
[0002] In clinical radiology, magnetic resonance imaging (MRI) offers unique advantages for disease diagnosis. Its multi-parameter, multi-sequence scans (T2-weighted imaging (T2WI), diffusion-weighted imaging (Diffusion Weighted Imaging (DWI), T1-weighted imaging (T1WI), and contrast-enhanced imaging) clearly demonstrate lesion morphology, blood supply, and other characteristics, displaying detailed lesion signs. Determining lesion type based on these signs is fundamental to clinical interpretability. Traditionally, lesion diagnosis has been based on clinical experience from images, but due to variability in manual identification, the efficiency and accuracy of lesion diagnosis cannot be guaranteed.
[0003] With the continuous development of artificial intelligence, especially deep learning, a growing number of disease-assisted diagnostics and applications have emerged. Automatic identification of MRI lesions using deep learning models will help improve the efficiency and accuracy of primary diagnosis. For example, common types of focal liver lesions include hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), metastasis, hemangioma, focal nodular hyperplasia (FNH), and cyst. Therefore, based on multi-parameter liver MRI images, machine models can be trained to automatically predict lesion types, thereby improving radiologists' work efficiency and reporting accuracy. The deep learning model used for automatic classification of liver MRI lesions takes lesion images from various MRI sequences as input and utilizes network layers such as convolutional neural networks (CNNs), fully connected layers (FCs), and non-linear activation functions to automatically extract lesion features and make classification predictions. In actual work, model predictions can be used as preliminary screening results or as additional testing and judgment for doctors' conclusions, and can be cross-verified with manual identification results to improve overall accuracy and reduce misdiagnosis rates.
[0004] However, existing research on automated MRI-based lesion diagnosis often relies on only a subset of MRI scan sequences or merely provides predictions of lesion type. In clinical radiology practice, lesion imaging features are crucial, serving as diagnostic evidence and providing interpretability for clinical judgment. However, due to the diverse imaging features of actual lesions (including at least 40 types), accurate and comprehensive predictions using classification models are difficult. Furthermore, common deep learning classification models (such as VGG, ResNet, and DenseNet) cannot effectively process inputs from a large number of modalities or predict causally related signal patterns. Different MRI sequences of lesions often contain both independent and coupled information, and the dependencies between predicted outputs vary. However, existing common deep learning network models are unable to separately process input modalities and output signals. Due to their "black box" operation and lack of interpretability of the model structure, they struggle to ensure high accuracy and reliability of the output results. Summary of the Invention
[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for classifying and identifying lesions in magnetic resonance images based on a causal circuit network model, which can improve the accuracy and interpretability of lesion type identification.
[0006] The object of the present invention can be achieved by the following technical solution: A method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model, comprising the following steps:
[0007] S1. Collect magnetic resonance multi-parameter sequence data, select sample data from it, perform preprocessing, and construct a data set;
[0008] S2. Build a causal circuit network model and use the data set to train the causal circuit network model to obtain a lesion recognition model;
[0009] S3. Preprocess the magnetic resonance imaging sequence to be identified, then input the lesion recognition model to obtain the corresponding lesion type and sign signal.
[0010] Furthermore, the step S1 includes the following steps:
[0011] S11, collecting scan images of magnetic resonance sequences, locating the center of the lesion, and aligning the lesions in different sequences;
[0012] S12. Annotate the scanned images of the magnetic resonance imaging sequence in combination with the corresponding pathological results;
[0013] S13, extracting the lesion ROI (region of interest) from the scanned image of the annotated magnetic resonance imaging sequence and performing data amplification processing;
[0014] S14. Use a five-fold cross-validation method to randomly divide the dataset into training, validation, and test sets.
[0015] Furthermore, the lesion ROI includes T1-weighted imaging, T2-weighted imaging, arterial phase AP, portal venous phase PVP, delayed phase DP, in-phase phase IP, anti-phase phase OP, diffusion enhancement weighted imaging DWI, and apparent diffusion coefficient imaging ADC.
[0016] Furthermore, the specific process of step S13 is: intercepting the same physical size in all magnetic resonance sequences and performing normalization processing, converting to obtain multiple lesion ROIs with the same pixel resolution, and performing data amplification processing on the lesion ROI, including multi-scale magnification and reduction, translation, random patching, rotation, flipping, and adding Gaussian noise.
[0017] Furthermore, the causal circuit network model in step S2 includes an image encoder, a causal circuit module, a signal encoder, a self-attention module, and a causal relay module connected in sequence, wherein the image encoder is used to extract lesion features of each modality image from the input magnetic resonance imaging sequence scan image lesion ROI;
[0018] The signal encoder is used to perform feature encoding on image signals of different modalities for signal feature learning;
[0019] The causal circuit module performs signal grooming according to the dependency image relationship of each prediction signal;
[0020] The self-attention module is used to integrate various signals globally, update the signal feature vector, and optimize the possible combinations of signals in the output;
[0021] The causal relay module is used to predict the sign signal and combine all the predicted signal vectors to form the overall characteristics of the lesion.
[0022] Furthermore, the image encoder and signal encoder are both composed of independent short-circuit convolutional network Res-Conv modules, which use the convolutional network layers to identify local visual features and avoid the gradient disappearance phenomenon during reverse feedback through short-circuit connections.
[0023] Furthermore, the causal relay module uses a fully connected layer to predict sign signals, including T1-weighted imaging, T2-weighted imaging, diffusion-enhanced weighted imaging DWI, apparent diffusion coefficient imaging ADC, arterial phase AP, portal venous phase PVP, delayed phase DP, progressive enhancement PROGRESS, central enhancement CENTER, fat signal FAT, and internal mosaicism INNER.
[0024] Furthermore, the causal relay module specifically processes each sign signal feature through a Sigmoid function and then combines it with the type signal to serve as the overall feature of the lesion.
[0025] Furthermore, the process of training the causal circuit network model using the data set in step S2 includes: setting training parameters;
[0026] The causal circuit network model is trained using the training set. After each epoch of training, the average feature accuracy is obtained on the validation set, and the model weight with the best accuracy is saved.
[0027] After training the model five times, five optimal model weights are obtained, and the ensemble model of the five models is used as the final model, that is, the lesion recognition model is obtained.
[0028] Furthermore, the training parameters include: batch size, learning rate, number of training epochs, loss function is cross-entropy function, and training optimizer is stochastic gradient descent (SGD).
[0029] Compared with the prior art, the present invention has the following advantages:
[0030] The present invention collects multi-parameter magnetic resonance imaging sequence data to construct a dataset. A causal circuit network model is then built and trained using the dataset to generate a lesion recognition model, which outputs corresponding lesion types and signs. Preprocessed multi-sequence magnetic resonance imaging is then fed into the lesion recognition model, where the causal circuit network automatically extracts, decouples, fuses, and optimizes the lesion types and signs, ultimately predicting the lesion type and signs. This improves the accuracy and interpretability of lesion type recognition.
[0031] The present invention designs a causal circuit network model, comprising an image encoder, a causal circuit module, a signal encoder, a self-attention module, and a causal relay module, connected in sequence. The image encoder and signal encoder are used to extract visual features. The causal circuit module is designed to guide information flow based on dependencies between features. A self-attention mechanism is used to perform global information fusion optimization and output a prediction combination. A causal relay model is used to predict sign signals, and all predicted signal vectors are combined to form the overall lesion feature. This allows for targeted extraction of lesion information for each magnetic resonance imaging (MRI) sequence. Through information decoupling and optimization, the optimal signal representation is obtained, and the predicted output combination is made consistent with clinical statistical characteristics, achieving model interpretability and high prediction accuracy.
[0032] The present invention performs data amplification processing on the lesion ROI, including multi-scale magnification and reduction, translation, random patching, rotation, flipping, and adding Gaussian noise, which can ensure that subsequent lesions of different sizes and shapes can be fully trained, thereby improving the reliability of the lesion recognition model.
[0033] In the present invention, the lesion recognition model can predict all the sign signals of the lesion (T1, T2, AP, PVP, DP, CENT, PROG, FAT, INNER, DWI, ADC) and the lesion type (Type). That is, it can not only identify and output the lesion type, but also predict the key sign signals of the lesion, thereby providing comprehensive information support for clinical decision-making. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 Schematic diagram of the method flow of the present invention;
[0035] Figure 2 This is a schematic diagram of the causal circuit network model;
[0036] Figure 3 This is a schematic diagram of the causal circuit network model structure;
[0037] Figure 4 Schematic diagram of short-circuit convolutional layer;
[0038] Figure 5 It is a schematic diagram of the overall characteristics of the lesion;
[0039] Figure 6 Schematic diagram of the model training process in the embodiment;
[0040] Figure 7 This is a schematic diagram of the model application deployment process in the embodiment;
[0041] Figure 8 Schematic diagram of the data preprocessing process in the embodiment. DETAILED DESCRIPTION
[0042] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0043] Example
[0044] This solution takes into account the extensive research in artificial intelligence technologies and deep learning models in the field of automated medical image analysis, which can provide technical support for automated auxiliary diagnosis of diseases. Convolutional neural networks have been widely applied in various computer vision tasks, achieving performance close to or exceeding that of human physicians in organ segmentation, lesion detection, and lesion classification. Transformer technology can calculate long-range token associations and is widely used in language analysis, multimodal prediction, and generative models. Therefore, this solution is based on commonly used deep learning modules (convolutional neural networks, Transformer, etc.), combined with the specific characteristics and goals of lesion classification in multi-series magnetic resonance imaging, to carry out targeted model structure design to achieve optimal prediction accuracy and model interpretability. The underlying principles it relies on include convolutional neural networks, deep neural networks, model training, back propagation, loss functions, stochastic gradient descent (SGD), and automatic updating of neural network parameters.
[0045] This paper proposes a method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model. Figure 1 As shown, the following steps are included:
[0046] S1. Collect magnetic resonance multi-parameter sequence data, select sample data from it, perform preprocessing, and construct a data set;
[0047] S2. Build a causal circuit network model and use the data set to train the causal circuit network model to obtain a lesion recognition model;
[0048] S3. Preprocess the magnetic resonance imaging sequence to be identified, then input the lesion recognition model to obtain the corresponding lesion type and sign signal.
[0049] The first purpose of this solution is to design a deep learning model that can process multiple sequences of magnetic resonance imaging of organ lesions, extract lesion features, and automatically predict lesion types and signs, to assist radiologists in diagnosis and increase efficiency; the second purpose is to design a special network connection, information extraction and fusion model for lesion sign signals and magnetic resonance scan images, in order to improve prediction accuracy and clinical interpretability. The causal circuit network model proposed in this solution is as follows: Figure 2 As shown in the figure, the causal circuit network can extract, decouple, fuse and optimize the coupled information in the input modalities, and ultimately predict the detailed sign characteristics of the lesion, thereby providing a reliable basis for clinical decision-making.
[0050] This embodiment applies the above scheme, taking the automatic classification of lesions in liver magnetic resonance images as an example. Its input is 10 3D magnetic resonance MR images, and its output is 11 lesion sign signals and lesion type signals. The causal circuit network can extract the lesion information of each MR sequence in a targeted manner, and obtain the best signal representation through information decoupling and information optimization, so that the predicted output combination conforms to the clinical statistical characteristics. The training of this deep learning relies on a specially collected, preprocessed, and labeled multi-sequence magnetic resonance image dataset of lesions. On the training dataset, the model calculates the loss function and updates the parameters through backpropagation, selects the best performing model on the validation dataset, and obtains performance results on the test dataset.
[0051] The model inputs are Region of Interest (ROIs) extracted from various MRI sequences of liver lesions, including T1-weighted imaging, T2-weighted imaging, arterial phase (AP), portal venous phase (PVP), delayed phase (DP), in-phase phase (IP), out-of-phase phase (OP), diffusion-enhanced weighted imaging (DWI with low b-value and DWI with high b-value), and apparent diffusion coefficient (ADC). The model outputs predictions of all lesion signs and lesion types, including T1, T2, DWI, ADC, AP, PVP, DP, PROGRESS, CENTER, FAT, INNER, and TYPE signals.
[0052] The model utilizes deep learning modules to process multimodal images, designing network connections based on task and image characteristics to achieve structural interpretability and high prediction accuracy. A convolutional encoder is used to extract visual features, and the network structure is designed to integrate dependencies to guide information flow between features. A self-attention mechanism is used to optimize global information fusion and output a combined prediction. The Transformer self-attention mechanism optimizes feature signatures and filters out low-probability predictions.
[0053] In this embodiment, the structure of the causal circuit network model is as follows: Figure 3 As shown in the figure, taking the classification of lesions in liver magnetic resonance images as an example, the leftmost part is the image capture area ROI of the lesion in each MR sequence. After the image encoder, causal circuit module, signal encoder, self-attention module, causal relay module, etc., all the sign signals (T1, T2, AP, PVP, DP, CENT, PROG, FAT, INNER, DWI, ADC) and type of the lesion are predicted.
[0054] Main structure and function:
[0055] 1. Image encoder: extracts lesion features from images of various modalities;
[0056] 2. Signal encoder: Encode features of different modal image signals for signal feature learning;
[0057] 3. Causal circuit module: performs signal grooming based on the dependency image relationship of each prediction signal;
[0058] 4. Self-attention module: This module is implemented using the classic QKV Transformer architecture. It globally integrates various signals, updates the signal feature vectors, and optimizes the possible combinations of signals in the output, making the results more consistent.
[0059] 5. Causal relay module: The symptom signal is predicted using a fully connected layer, and all predicted signal vectors are combined to form the overall characteristics of the lesion.
[0060] The image encoder and signal encoder are composed of independent short-circuit convolutional network (Res-Conv) modules. Each Res-Conv module is composed of multiple layers of short-circuit convolutional layers, such as Figure 4 As shown in the figure, convolutional layers can effectively identify local visual features, while short-circuit connections can prevent gradient vanishing during reverse feedback. Regularization layers and ReLu nonlinear layers can increase network capacity and stability.
[0061] In the causal relay module, each sign signal feature is combined with the Type signal after the Sigmoid function to form the overall lesion signal to predict the lesion type. Figure 5 The lesion's overall joint feature vector has clear physical meanings, providing a more accurate foundation for comparative analysis of the feature spaces of different lesions.
[0062] The application process of this embodiment is mainly divided into two parts: model training and application deployment.
[0063] Model training process: prepare training data, cross-group data in five-fold, build GPU training platform, set training parameters, conduct training and select the best model. The specific process is as follows: Figure 6 shown.
[0064] Application deployment process: prepare the workstation to deploy the model, monitor the local hard disk or network address, extract the case images in the request and pre-process them, perform multi-model prediction and integration, the specific process is as follows Figure 7 shown.
[0065] Specifically:
[0066] 1. Data collection: Taking liver MRI scans as an example, scan images of all MRI sequences are collected, the center of the lesion is located through manual or automatic detection models, and the lesions in different sequences are aligned.
[0067] 2. Use lesion signs as a prediction task: Professional radiologists will annotate them in combination with pathological results as the gold standard (Ground Truth).
[0068] 3. Data ROI extraction: Figure 8 As shown in the figure, for all sequences, the same physical size (8 cm*8 cm*8 cm) was captured, normalized (grayscale and resolution), and converted into ROI with a resolution of 80*80*8 pixels.
[0069] 4. Data augmentation of lesion ROI: This includes multi-scale magnification and reduction, translation, random patching, rotation, flipping, and adding Gaussian noise to ensure that lesions of different sizes and shapes can be fully trained.
[0070] 5. Data partitioning: This embodiment adopts a five-fold cross-validation method and randomly divides the data set into a training set (64%), a validation set (16%), and a test set (20%).
[0071] 6. Computing hardware preparation: at least a 24-core CPU workstation, at least 64GB of memory, and at least a 24GB NVIDIA GPU.
[0072] 7. Set the model training parameters: batch size to 128, learning rate to 0.01, training for 200 epochs, cross-entropy as the loss function, and stochastic gradient descent (SGD) as the optimizer.
[0073] 8. Model preservation: Train the causal circuit network model (CRN). After each epoch of training, obtain the average feature accuracy on the validation set and save the model weight with the best accuracy.
[0074] 9. Final model: Using five-fold cross validation, after training the model five times, we get the five best model weights, and use the ensemble model of the five models as the final model.
[0075] 10. Model performance testing: The integrated model obtained above is used for testing in external independent data to obtain accurate classification of lesion signs and lesions.
[0076] To verify the effectiveness of this solution, this embodiment also performs an interpretive analysis on the prediction results. The steps of the "lesion sign-type interpretability analysis" predicted by the model are as follows:
[0077] 1) Grouping and counting lesion sign signals according to correlation, merging positively correlated (i.e., coupled) sign signals and juxtaposing negatively correlated (i.e., mutually exclusive) ones;
[0078] 2) Calculate and outline the causal relationship diagram of "signal signal-lesion type" and calculate the probability of the inference path from different sign signals to lesion type;
[0079] 3) Compare the causal relationship diagrams in the model predictions with those in the gold standard, and derive the consistency and accuracy of the prediction path based on the connection relationship and probability of the diagrams. In addition, eliminate impossible sign combinations and issue warnings for low-probability prediction signal combinations.
[0080] In summary, this solution enables more comprehensive lesion characteristic prediction, improved clinical interpretability, causal interpretability of the model structure, and greater consistency in predicted outputs. Specifically, this solution not only outputs lesion type but also predicts key lesion signs, providing comprehensive information support for clinical decision-making. Furthermore, the relationship between sign combinations and sign-type probability analysis provide a basis for model predictions, making the outputs interpretable. By designing independent feature encoders and a dependency-based structural combination, this solution achieves better training and fitting results while also improving the transparency and interpretability of the model structure. This solution uses a self-attention mechanism to update the representation of all predicted signals, optimizing the overall lesion representation. This can help reduce anomalous representations in the feature space and achieve greater consistency in the output combination. When applied in practice, this solution can predict lesion signs to assist in accurate lesion type determination and provide interpretability for clinical decision-making. The causal association module incorporated into the model structure clarifies the dependency between input modalities and output signals, optimizing the model structure. Furthermore, it can automatically optimize the output combination of predicted signals, improving recognition and prediction accuracy.
Claims
1. A method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model, characterized in that: The following steps are involved: S1. Collect magnetic resonance multi-parameter sequence data, select sample data from it, perform preprocessing, and construct a data set; S2. Build a causal circuit network model and use the data set to train the causal circuit network model to obtain a lesion recognition model; S3. Preprocess the magnetic resonance imaging sequence to be identified, then input the lesion recognition model to obtain the corresponding lesion type and sign signal.
2. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 1, characterized in that: The step S1 comprises the following steps: S11, collecting scan images of magnetic resonance sequences, locating the center of the lesion, and aligning the lesions in different sequences; S12. Annotate the scanned images of the magnetic resonance imaging sequence in combination with the corresponding pathological results; S13, extracting the lesion ROI from the scanned image of the annotated magnetic resonance sequence and performing data amplification processing; S14. Use a five-fold cross-validation method to randomly divide the dataset into training, validation, and test sets.
3. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 2, characterized in that: The lesion ROI includes T1-weighted imaging, T2-weighted imaging, arterial phase AP, portal venous phase PVP, delayed phase DP, in-phase phase IP, anti-phase phase OP, diffusion enhancement weighted imaging DWI, and apparent diffusion coefficient imaging ADC.
4. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 2, wherein: The specific process of step S13 is as follows: intercepting the same physical size in all magnetic resonance sequences and performing normalization processing, converting to obtain multiple lesion ROIs with the same pixel resolution, and performing data amplification processing on the lesion ROI, including multi-scale magnification and reduction, translation, random patching, rotation, flipping, and adding Gaussian noise.
5. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 2, wherein: The causal circuit network model in step S2 includes an image encoder, a causal circuit module, a signal encoder, a self-attention module, and a causal relay module connected in sequence, wherein the image encoder is used to extract lesion features of each modality image from the input magnetic resonance imaging sequence scan image lesion ROI; The signal encoder is used to perform feature encoding on image signals of different modalities for signal feature learning; The causal circuit module performs signal grooming according to the dependency image relationship of each prediction signal; The self-attention module is used to integrate various signals globally, update the signal feature vector, and optimize the possible combinations of signals in the output; The causal relay module is used to predict the sign signal and combine all the predicted signal vectors to form the overall characteristics of the lesion.
6. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 5, characterized in that: The image encoder and signal encoder are both composed of independent short-circuit convolutional network Res-Conv modules, which use the convolutional network layers to identify local visual features and avoid the gradient disappearance phenomenon during reverse feedback through short-circuit connections.
7. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 6, characterized in that: The causal relay module uses a fully connected layer to predict sign signals, including T1-weighted imaging, T2-weighted imaging, diffusion-enhanced weighted imaging DWI, apparent diffusion coefficient imaging ADC, arterial phase AP, portal venous phase PVP, delayed phase DP, progressive enhancement PROGRESS, central enhancement CENTER, fat signal FAT, and internal mosaicism INNER.
8. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 6, characterized in that: The causal relay module specifically passes each sign signal feature through the Sigmoid function and then combines it with the type signal to serve as the overall feature of the lesion.
9. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 2, wherein: The process of training the causal circuit network model using the data set in step S2 includes: setting training parameters; The causal circuit network model is trained using the training set. After each epoch of training, the average feature accuracy is obtained on the validation set, and the model weight with the best accuracy is saved. After training the model five times, five optimal model weights are obtained, and the integrated model of the five models is used as the final model, that is, the lesion recognition model is obtained.
10. The method for classifying and identifying lesions in magnetic resonance imaging based on a causal circuit network model according to claim 9, characterized in that: The training parameters include: batch size, learning rate, number of training epochs, loss function is cross entropy function, and training optimizer is stochastic gradient descent.