A method, system, device and terminal for classifying orbital lymphoma and inflammatory pseudotumor

Abstract

The application belongs to the technical field of medical image processing and computer-aided diagnosis, and discloses an orbital lymphoma and inflammatory pseudotumor classification method, system, device and terminal, in the obtained orbital DCE-MRI image, the tumor region and the eye cone triangular region in the original image are manually segmented and pretreated; the neural network is used for extracting the features of the tumor region, and the features are subjected to clustering statistical analysis as the tumor region features; the eye cone triangular region is subjected to feature extraction as a similarity judgment standard, and the self-classification and self-recovery network is used for extracting the eye cone region features; the original image is used as the input network and combined with the features of the tumor region and the eye cone triangular region to train the network model; the multi-modal orbital data is input into the classification model for processing, so that the orbital lymphoma and inflammatory pseudotumor are classified. The application comprehensively considers the texture features of the tumor and the depth features of the eye cone region, effectively improves the prediction accuracy, and has the characteristics of high accuracy.

Description

Technical Field

[0001] This invention belongs to the fields of medical image processing and computer-aided diagnosis technology, and in particular relates to a method, system, device and terminal for classifying orbital lymphoma and inflammatory pseudotumor. Background Technology

[0002] Currently, orbital lymphoma is the most common malignant tumor of the orbit in adults, and is usually treated with local radiotherapy. Inflammatory pseudotumor is an idiopathic proliferative inflammatory disease, which is usually treated with anti-inflammatory drugs such as corticosteroids and adjuvant antibiotics. The clinical imaging information of orbital lymphoma and inflammatory pseudotumor often overlaps, making them difficult to distinguish using routine histology. Biopsy provides the gold standard for pathological examination of orbital lymphoma and inflammatory pseudotumor, but it is invasive and can cause complications. Preoperative magnetic resonance imaging (MRI) provides an ideal means for the non-invasive diagnosis of orbital diseases. Compared with biopsy, preoperative MRI is less invasive, but the physician's experience and the hospital's imaging equipment can affect the diagnosis. Therefore, there is an urgent need for a non-invasive, accurate, and efficient classification method for orbital lymphoma and inflammatory pseudotumor. While classification and diagnostic systems for orbital lymphoma and inflammatory pseudotumors have been proposed, most employ radiomics approaches. This requires complex analytical steps, including manual feature extraction, feature selection, and predictive model construction. The inherent randomness of each step can lead to inconsistent results. Few studies have used deep learning to differentiate between orbital lymphoma and inflammatory pseudotumors, focusing primarily on information from the orbital triangle while neglecting tumor information. Currently, there are no studies combining research on the orbital triangle and tumor regions.

[0003] Currently, radiomics-based analysis involves complex steps, each with inherent randomness, leading to inconsistent results. Deep learning methods are less studied in this area and focus only on partial information, failing to comprehensively consider the shared information between the tumor and the orbital cone triangle. However, due to the randomness of the steps, effective feature extraction and classification are challenging in diagnosing orbital lymphoma and inflammatory pseudotumors using radiomics-based methods. Research on deep learning in orbital image classification is still in its early stages, and constructing an efficient model for classifying orbital lymphoma and inflammatory pseudotumors remains a significant challenge.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0005] (1) Most existing classification and diagnostic systems for orbital lymphoma and inflammatory pseudotumor adopt radiomics approach, which involves complex analysis steps and inconsistent classification results due to the randomness of each step.

[0006] (2) Currently, there are few methods that use deep learning to distinguish between orbital lymphoma and inflammatory pseudotumor, and these methods only focus on information about the orbital triangle or only consider the tumor area, without comprehensively considering the common information of the tumor and the orbital triangle.

[0007] (3) There are currently no reports on the combined analysis techniques for the ocular cone triangle and tumor regions. Summary of the Invention

[0008] To address the problems existing in the prior art, the present invention provides a method, system, device and terminal for classifying orbital lymphoma and inflammatory pseudotumor, and particularly relates to a method, system, device and terminal for classifying orbital lymphoma and inflammatory pseudotumor based on multidimensional information fusion.

[0009] This invention is implemented as follows: a classification method for orbital lymphoma and inflammatory pseudotumor. The method includes: manually segmenting and preprocessing the tumor region and the ocular cone triangle region in the acquired orbital dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) image; using the preprocessed 3D image as a sample to divide the sample dataset; extracting features from the tumor region using a neural network and performing clustering statistical analysis on the features as tumor region features; extracting features from the ocular cone triangle region as a similarity evaluation criterion, using a self-classification and self-recovery network to extract ocular cone region features; inputting the training dataset into the network and training the network model by combining the features of the tumor region and the ocular cone triangle region, and using test set data for prediction; inputting multimodal data into the classification model for processing to achieve the classification task of orbital lymphoma and inflammatory pseudotumor.

[0010] Furthermore, the classification method for orbital lymphoma and inflammatory pseudotumor includes the following steps:

[0011] Step 1, Segmentation and Processing of Original Images: In the acquired orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original images are manually segmented using the Medical Image Interaction Tool Kit (MITK) software; the raw format data is processed into mat format data; the region of interest with the most values ​​greater than 0 is selected as the largest slice, and five slices are selected sequentially; the processed 3D original image, the 3D image of the tumor region, and the 3D image of the ocular cone triangle region are used as sample datasets, and training and testing sets are divided into them.

[0012] Step 2, Feature extraction of the tumor region: Extract 9×9 image patches centered on each pixel of the segmented tumor region, and stack the extracted patches in the third dimension; build a two-layer neural network consisting of encoding and decoding parts, and train the neural network using the training dataset; input the training set and test set into the neural network to automatically extract features; use the k-means++ clustering algorithm to cluster the features into k classes; perform statistical analysis on the clustered features to obtain the tumor region features;

[0013] Step 3, Feature extraction of the eye cone triangle region: Train an autoencoder network to extract deep features of the image as similarity to rank the data; generate five position coordinates and generate visual words at the five coordinate points; calculate the contrast loss function between the original image and the restored image through the self-recovery branch, and train a five-classification network; select the optimal model to obtain the deep feature extraction model of the eye cone triangle region, and extract the depth features of the eye cone region.

[0014] Step 4, fusion feature classification based on ViT (Vision Transformer) model: Input the training data of the original image into the pre-trained ViT model, combine the feature data obtained by the Transformer encoder with the texture features of the tumor and the depth features of the eye cone triangle, and classify orbital lymphoma and inflammatory pseudotumor through the multilayer perceptron (MLP) classification head.

[0015] Step 5, Multimodal Classification: The multimodal data is processed, and a soft voting strategy is used to integrate the modalities, ultimately completing the classification task for orbital lymphoma and inflammatory pseudotumor.

[0016] Furthermore, in step one, the segmentation and processing of the original image in the acquired orbital DCE-MRI image includes:

[0017] (1) In the obtained orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original image were manually segmented using MITK software;

[0018] (2) Convert the raw format data of the original image, the segmented tumor region and the eye cone triangle region into mat format data;

[0019] (3) Count the number of values ​​greater than 0 in the image and select the region of interest with the most values ​​as the largest slice. Select two slices above and below the largest slice, for a total of five slices.

[0020] (4) The processed original three-dimensional image, the three-dimensional image of the tumor region and the three-dimensional image of the eye cone triangle are used as sample datasets, and the training set and test set are randomly divided.

[0021] Furthermore, the feature extraction of the tumor region in step two includes:

[0022] (1) Extract a 9×9 image patch block centered on each pixel of the segmented tumor region, and superimpose the extracted patch blocks in the third dimension;

[0023] (2) Construct a two-layer neural network consisting of an encoding and decoding part, and train the neural network using a training dataset;

[0024] (3) Input the training set and test set into the trained neural network to automatically extract features;

[0025] (4) Use the k-means++ clustering algorithm to cluster the features into k classes;

[0026] (5) Statistical analysis of the clustered features yields the tumor region characteristics.

[0027] Furthermore, the feature extraction of the eye cone triangle region in step three includes:

[0028] (1) Use the training set data to train the autoencoder network, and extract the deep features of the images as similarity to rank the data;

[0029] (2) Generate five location coordinates and generate visual words at the five coordinate points;

[0030] (3) Calculate the contrast loss function between the original image and the restored image through the self-recovery branch, and train the five-class classification network; wherein, the training is performed by minimizing the standard classification cross-entropy loss function, as shown in the following formula:

[0031]

[0032] In the formula, L cls Let B represent the classification loss function; C represent the batch size; y represent the number of visual words; and P represent the ground truth and network prediction, respectively. bc P represents the true value when the batch size is b and the number of visual words is c. bc This represents the predicted value when the batch size is b and the number of visual words is c, where b and c represent values ​​from 1 to B and from 1 to C, respectively.

[0033] The recovery network is trained by minimizing the L2 distance between the original visual word and the reconstructed visual word, as shown in the following formula:

[0034]

[0035] In the formula, L rec Let B represent the loss function of the recovery network, B represent the batch size, and x represent the loss function of the recovery network. iand x i 'These represent the original visual word and the restored visual word, respectively;

[0036] (4) Select the optimal model to obtain the deep feature extraction model of the eye cone triangle region;

[0037] (5) Extract the depth features of the eye cone region.

[0038] Furthermore, step four, which involves classification based on the ViT model, includes:

[0039] (1) Input the training data of the original image into the pre-trained ViT model;

[0040] (2) The feature data obtained after passing through the Transformer encoder is combined with the texture features of the tumor and the depth features of the ocular cone triangle.

[0041] (3) Classification of orbital lymphoma and inflammatory pseudotumor by MLP classification head.

[0042] Furthermore, step five involves processing the multimodal data to classify orbital lymphomas and inflammatory pseudotumors, including:

[0043] (1) Perform data processing, tumor region feature extraction, eye cone region extraction, and ViT network feature fusion steps on the three modal data T1WI, T1WI+C and T2WI respectively;

[0044] (2) The results are classified into orbital lymphoma and inflammatory pseudotumor by using a soft voting multimodal integration strategy and the results are classified according to the probability statistics of the division.

[0045] Another object of the present invention is to provide an orbital lymphoma and inflammatory pseudotumor classification system applying the aforementioned orbital lymphoma and inflammatory pseudotumor classification method, the orbital lymphoma and inflammatory pseudotumor classification system comprising:

[0046] The data acquisition module is used to acquire standard-compliant orbital DCE-MRI data;

[0047] The data preprocessing module is used to manually segment the tumor region and the ocular cone triangle region in the acquired orbital DCE-MRI images using MITK software.

[0048] The tumor region feature extraction module is used to extract a 9×9 image patch centered on each pixel of the segmented tumor region and automatically extract tumor region features to express tumor texture information.

[0049] The eye cone triangle region feature extraction module is used to train the model with the eye cone triangle region image as input and to extract the depth features of the eye cone region using the eye cone triangle region deep feature extraction model.

[0050] The ViT fusion feature classification module is used to input the training data of the original image into the pre-trained ViT model and fuse the feature data with the texture features of the tumor and the depth features of the ocular cone triangle.

[0051] The multimodal classification module is used to process multimodal data, integrate modalities using a soft voting strategy, and complete the classification task of orbital lymphoma and inflammatory pseudotumor.

[0052] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for classifying orbital lymphoma and inflammatory pseudotumor.

[0053] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for classifying orbital lymphoma and inflammatory pseudotumor.

[0054] Another objective of this invention is to provide an information data processing terminal for implementing the aforementioned orbital lymphoma and inflammatory pseudotumor classification system.

[0055] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0056] First, addressing the technical problems existing in the prior art and the difficulty of solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:

[0057] This invention automatically extracts features, eliminating the need for manual feature selection. While improving accuracy, it also effectively reduces model complexity and randomness, resulting in a stronger generalization ability. The effective deep learning network model can be used for the classification of orbital tumors and inflammatory pseudotumors, providing an auxiliary diagnostic and treatment method for orbital diseases. It avoids the harm to the human body caused by biopsies and effectively reduces errors caused by manual diagnosis.

[0058] This invention fully utilizes multimodal data from orbital DCE-MRI, combining features from the tumor region and the ocular cone triangle region, and processing the data using the ViT framework to comprehensively consider both global and local information of the orbit. It replaces manual feature extraction with neural network extraction, avoiding randomness in the results and improving the model's generalization ability. Furthermore, this invention can be used for the classification of orbital lymphomas and inflammatory pseudotumors, providing auxiliary diagnostic and treatment tools for clinicians and offering diagnostic references for physicians.

[0059] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:

[0060] This invention provides a non-invasive diagnostic method for orbital lymphoma and inflammatory pseudotumor, which solves the problem of invasive damage caused by biopsy. It comprehensively considers the texture features of the tumor, the depth features of the orbital cone region, and the multimodal information of the orbit, effectively improving the prediction accuracy and model stability. It has the characteristics of requiring only a small amount of labeled data, high accuracy, and saving data processing costs.

[0061] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:

[0062] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0063] This invention can be used for the classification and diagnosis of orbital lymphoma and inflammatory pseudotumor in clinical practice, providing a non-invasive diagnostic method for clinical medicine, which can assist doctors in making a faster and more accurate diagnosis of the disease. Compared with existing classification methods for orbital lymphoma and inflammatory pseudotumor, this invention provides a classification method based on multidimensional information fusion, which reduces the time cost required for classification by the original method, improves the efficiency of classification, and effectively improves the classification accuracy. This invention can be used for clinical applications in hospitals and product development.

[0064] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0065] Existing classification methods for orbital lymphoma and inflammatory pseudotumors mostly use radiomics methods, which involve complex data processing steps with high randomness. Only a very small number of researchers have used deep learning methods for classification studies. However, deep learning methods only focus on the features of the orbital triangle or only on the tumor region, without considering the combined features of the orbital triangle and the tumor region. This invention considers the features of the orbital triangle and the tumor region together and combines the multimodal information of the data to provide a classification method based on multidimensional information fusion.

[0066] (3) The technical solution of the present invention overcomes technical bias:

[0067] Existing research methods focus only on information from the orbital triangle or tumor region, neglecting data integrity and overlooking tumor texture or orbital triangle depth. This invention considers the features of both the orbital triangle and tumor region together, incorporating multimodal data to create a more holistic and comprehensive model, thus overcoming the technical biases of previous solutions. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a flowchart of the classification method for orbital lymphoma and inflammatory pseudotumor provided in an embodiment of the present invention;

[0070] Figure 2 This is a diagram of the neural network model structure for tumor region feature extraction provided in an embodiment of the present invention;

[0071] Figure 3 This is a model structure diagram for feature extraction of the eye cone triangle region provided in an embodiment of the present invention;

[0072] Figure 4 This is a structural diagram of the orbital multimodal data processing model provided in an embodiment of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0074] To address the problems existing in the prior art, the present invention provides a method, system, device, and terminal for classifying orbital lymphoma and inflammatory pseudotumor. The present invention will be described in detail below with reference to the accompanying drawings.

[0075] like Figure 1 As shown, the classification method for orbital lymphoma and inflammatory pseudotumor provided in this embodiment of the invention includes the following steps:

[0076] S101, In the acquired orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original image are manually segmented and preprocessed, and the three-dimensional image is used as a sample;

[0077] S102, use a neural network to extract features from the segmented tumor region, use the k-means++ clustering algorithm to cluster the features, and use the statistical analysis as the tumor region features;

[0078] S103, train an autoencoder network to extract features from the eye cone triangle region as a similarity criterion, and use a self-classification and self-recovery network to extract depth features of the eye cone region;

[0079] S104 uses the original image as input to train the network, fusing features from the tumor region and the ocular cone triangle region; it inputs multimodal data into the classification model to classify orbital lymphoma and inflammatory pseudotumor.

[0080] As a preferred embodiment, the method for classifying orbital lymphoma and inflammatory pseudotumor provided in this invention specifically includes the following steps:

[0081] (1) Segmentation and processing of the original image: In the obtained orbital DCE-MRI image, the tumor region and the triangular region of the eye cone in the original image were manually segmented using MITK software; the raw format data was processed into mat format data; the region of interest with the most values ​​greater than 0 was selected as the largest slice, and five slices were selected in sequence; the processed three-dimensional original image, the three-dimensional image of the tumor region and the three-dimensional image of the triangular region of the eye cone were used as sample datasets, and the training set and test set were divided.

[0082] The segmentation and processing of the original image in the acquired orbital DCE-MRI image includes:

[0083] (1.1) In the obtained orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original image were manually segmented using MITK software;

[0084] (1.2) Convert the raw format data of the original image, the segmented tumor region and the eye cone triangle region into mat format data;

[0085] (1.3) Count the number of values ​​greater than 0 in the image and select the region of interest with the most values ​​as the largest slice. Select two slices above and below the largest slice, for a total of five slices.

[0086] (1.4) The processed original 3D image, the 3D image of the tumor region and the 3D image of the eye cone triangle are used as sample datasets, and the training set and test set are randomly divided.

[0087] (2) Feature extraction of the tumor region: 9×9 image patch blocks are extracted with each pixel of the segmented tumor region as the center, and the extracted patch blocks are superimposed in the third dimension; a two-layer neural network is built, consisting of encoding and decoding parts, and the neural network is trained using the training dataset; the training set and test set are input into the neural network to automatically extract features; the k-means++ clustering algorithm is used to cluster the features into k classes; statistical analysis of the clustered features is performed to obtain the tumor region features.

[0088] The feature extraction of the tumor region is performed according to the following steps:

[0089] (2.1) Extract a 9×9 image patch block centered on each pixel of the segmented tumor region, and stack the extracted patch blocks in the third dimension;

[0090] (2.2) Construct a two-layer neural network, such as Figure 2 As shown, the system consists of encoding and decoding parts. The encoder consists of two convolutional layers and two max-pooling layers. The first convolutional layer has 4 kernels and a kernel size of 3, and the second convolutional layer has 8 kernels and a kernel size of 3. The stride of the max-pooling layer is 2. The decoder consists of three convolutional layers and two upsampling layers. The first convolutional layer has 8 kernels and a kernel size of 3, the second convolutional layer has 4 kernels and a kernel size of 3, and the third convolutional layer has 1 kernel and a kernel size of 3. The neural network is trained using a training dataset.

[0091] (2.3) Input the training set and test set into the trained neural network to automatically extract features;

[0092] (2.4) Use the k-means++ clustering algorithm to cluster the features into k classes;

[0093] (2.5) Statistical analysis of the clustered features yields the tumor region characteristics.

[0094] (3) Feature extraction of the eye cone triangle region, such as Figure 3 As shown, an autoencoder network is trained to extract deep features from images and use them as similarity scores to rank the data. Five location coordinates are generated, and visual words are generated at these five coordinate points. A self-recovery branch is used to calculate the contrast loss function between the original and recovered images, and a five-class classification network is trained. The optimal model is selected to obtain the deep feature extraction model for the eye cone triangle region. The depth features of the eye cone region are extracted. The feature extraction of the eye cone triangle region is performed according to the following steps:

[0095] (3.1) Train an autoencoder network and extract deep features of the images as similarity to rank the data;

[0096] (3.2) Generate five location coordinates and generate visual words at these five coordinate points;

[0097] (3.3) Calculate the contrast loss function between the original image and the restored image through the self-recovery branch, and train the five-class classification network;

[0098] (3.4) Select the optimal model to obtain the deep feature extraction model of the eye cone triangle region;

[0099] (3.5) Extract the depth features of the eye cone region.

[0100] (4) Classification is achieved based on the existing VIT model. The training data of the original image is input into the pre-trained ViT model. The feature data obtained after passing through the Transformer encoder is combined with the texture features of the tumor and the depth features of the eye cone triangle. Orbital lymphoma and inflammatory pseudotumor are classified by the MLP classification head.

[0101] The classification based on the existing VIT model is carried out according to the following steps:

[0102] (4.1) Input the training data of the original images into the pre-trained ViT model;

[0103] (4.2) The feature data obtained after passing through the Transformer encoder is combined with the texture features of the tumor and the depth features of the eye cone triangle;

[0104] (4.3) Orbital lymphoma and inflammatory pseudotumor were classified by the MLP classification head.

[0105] (5) Process the multimodal data, such as Figure 4 As shown, the method of this invention is used to classify orbital lymphoma and inflammatory pseudotumor. The three modalities are processed in the same way. The modalities are integrated using a soft voting strategy to finally complete the classification of orbital lymphoma and inflammatory pseudotumor.

[0106] The process of processing multimodal data to classify orbital lymphomas and inflammatory pseudotumors follows these steps:

[0107] (5.1) Perform data processing on the three modal data T1WI, T1WI+C and T2WI respectively, extract tumor region features, extract eye cone region, and fuse features using ViT network;

[0108] (5.2) The results are classified into orbital lymphoma and inflammatory pseudotumor by using a soft-voting multimodal integration strategy and the results are classified according to the probability statistics of the division.

[0109] The orbital lymphoma and inflammatory pseudotumor classification system provided in this invention includes:

[0110] The data acquisition module is used to acquire standard-compliant orbital DCE-MRI data;

[0111] The data preprocessing module is used to manually segment the tumor region and the ocular cone triangle region in the acquired orbital DCE-MRI images using MITK software.

[0112] The tumor region feature extraction module is used to extract a 9×9 image patch centered on each pixel of the segmented tumor region and automatically extract tumor region features to express tumor texture information.

[0113] The eye cone triangle region feature extraction module is used to train the model with the eye cone triangle region image as input and to extract the depth features of the eye cone region using the eye cone triangle region deep feature extraction model.

[0114] The ViT fusion feature classification module is used to input the training data of the original image into the pre-trained ViT model and fuse the feature data with the texture features of the tumor and the depth features of the ocular cone triangle.

[0115] The multimodal classification module is used to process multimodal data, integrate modalities using a soft voting strategy, and complete the classification task of orbital lymphoma and inflammatory pseudotumor.

[0116] To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides specific product or related technology application examples of the technical solution claimed.

[0117] This invention can be applied to a computer-aided diagnostic system for orbital lymphoma and inflammatory pseudotumors. The system receives DCE-MRI image data of the patient's eye as training input, then segments and preprocesses the data. The processed images are then fed into a classification model for training, resulting in a classification model that is stored within the computer-aided diagnostic system for orbital lymphoma and inflammatory pseudotumors. The trained system can then receive DCE-MRI image data of the patient's eye as test input, feeding the test data into the trained classification model to obtain prediction results. Finally, the computer-aided diagnostic system for orbital lymphoma and inflammatory pseudotumors outputs the classification and diagnostic results for orbital lymphoma and inflammatory pseudotumors.

[0118] As a preferred embodiment, the method for classifying orbital lymphoma and inflammatory pseudotumor provided in this invention specifically includes the following steps:

[0119] (1) Acquisition of orbital DCE-MRI image data, segmentation and processing of the original image, the specific process is as follows:

[0120] (1a) The data used in the embodiments of the present invention are orbital DCE-MRI images;

[0121] (1b) The inclusion criteria are as follows: 1) MRI examination within 7 days prior to biopsy or surgery; 2) pathologically confirmed primary orbital lymphoma or inflammatory pseudotumor; 3) patients with no history of orbital surgery or treatment prior to examination;

[0122] (1c) The exclusion criteria are as follows: 1) The number of lesion-containing slices in the MRI imaging is less than 5; 2) The image quality is poor, such as obvious motion or magnetic susceptibility artifacts.

[0123] (1d) A total of 93 patients were included, including 39 patients with orbital lymphoma and 54 patients with inflammatory pseudotumor. 74 of these patients were used as the training set and 19 patients were used as the test set.

[0124] (1e) In the obtained orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original image were manually segmented using MITK software;

[0125] (1f) Convert the raw format data of the original image, the segmented tumor region and the eye cone triangle region into mat format data.

[0126] (2) Manual segmentation and preprocessing of the orbital image, the specific process is as follows:

[0127] (2a) In the obtained orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original image were manually segmented using MITK software;

[0128] (2b) Convert the raw format data of the original image, the segmented tumor region and the eye cone triangle region into mat format data;

[0129] (2c) Count the number of values ​​greater than 0 in the image, select the region of interest with the most values ​​as the largest slice, and select two slices above and below the largest slice, for a total of five slices;

[0130] (2d) The processed 3D original image, the 3D image of the tumor region and the 3D image of the eye cone triangle are used as the sample dataset, and the training set and test set are randomly divided.

[0131] (3) Features are extracted from the tumor region using a neural network, and cluster statistical analysis is performed on the features to serve as tumor region features. The specific process is as follows:

[0132] (3a) Extract a 9×9 image patch block centered on each pixel of the segmented tumor region, and stack the extracted patch blocks in the third dimension;

[0133] (3b) Construct a two-layer neural network, such as Figure 2 As shown, the system consists of encoding and decoding parts. The encoder consists of two convolutional layers and two max-pooling layers. The first convolutional layer has 4 kernels and a kernel size of 3, and the second convolutional layer has 8 kernels and a kernel size of 3. The stride of the max-pooling layer is 2. The decoder consists of three convolutional layers and two upsampling layers. The first convolutional layer has 8 kernels and a kernel size of 3, the second convolutional layer has 4 kernels and a kernel size of 3, and the third convolutional layer has 1 kernel and a kernel size of 3. The neural network is trained using a training dataset.

[0134] (3c) Input the training set and test set into the trained neural network to automatically extract features;

[0135] (3d) Use the k-means++ clustering algorithm to cluster the features into 18 classes;

[0136] (3e) Count the number of clustered features belonging to each class to obtain 18 specific probabilities for each pixel patch, which are used as tumor features to obtain tumor region features.

[0137] (4) Feature extraction of the eye cone triangle region is used as a similarity criterion. Self-classification and self-recovery networks are employed to extract features from the eye cone region, such as... Figure 3 As shown, the specific process is as follows:

[0138] (4a) Use the training set data to train the autoencoder network, extract the latent spatial features of the images to learn the same mapping from all patients to themselves, and use the similarity to rank the data;

[0139] (4b) Randomly select a patient as a reference and form a set of K nearest neighbors for the entire training set based on the L2 distance in the latent space;

[0140] (4c) Select 5 sets of coordinates and specify unique pseudo-labels. Each pseudo-label corresponds to K patch instances.

[0141] (4d) After perturbing the 5 sets of patch instances, the data is input into the encoder. This part is trained by minimizing the standard classification cross-entropy loss function, as shown in the following formula:

[0142]

[0143] Where B represents the batch size; C represents the number of visual words; y and P represent the ground truth (a pseudo-label vector) and the network prediction, respectively;

[0144] (4e) Decoder decoding: This part trains the recovery network by minimizing the L2 distance between the original visual word and the reconstructed visual word, as shown in the following formula:

[0145]

[0146] Where B represents the batch size, and x and x' represent the original visual word and the restored visual word, respectively.

[0147] (5) Input the training dataset into the ViT network and train the network model by combining the features of the tumor region and the ocular cone triangle. Use the test set for prediction. The specific process is as follows:

[0148] (5a) Input the training data of the original image into the pre-trained ViT model;

[0149] (5b) The feature data obtained after passing through the Transformer encoder is combined with the texture features of the tumor and the depth features of the ocular cone triangle;

[0150] (5c) Classification of orbital lymphoma and inflammatory pseudotumor by MLP classification head.

[0151] (6) The multimodal data is processed to achieve the classification task of orbital lymphoma and inflammatory pseudotumor. The specific process is as follows:

[0152] (6a) Perform data processing on the three modal data T1WI, T1WI+C and T2WI respectively, extract tumor region features, extract eye cone region, and fuse features using ViT network;

[0153] (6b) The results are classified into orbital lymphoma and inflammatory pseudotumor by using a soft-voting multimodal integration strategy and the results are classified according to the probability statistics of the division.

[0154] The evaluation criteria for the system proposed in this embodiment include receiver operating characteristic (ROC) curves and area under the curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), and negative predictive value (NPV). These indicators are used to evaluate the classification performance of orbital lymphoma and inflammatory pseudotumor. The values ​​of these indicators are all between [0, 1], and the closer they are to 1, the better the model's predictive performance. In the test set, each individual modality was processed and the modalities were integrated to classify orbital lymphoma and inflammatory pseudotumor. The results are shown in Table 1.

[0155] Table 1. Classification results of orbital lymphoma and inflammatory pseudotumor

[0156]

[0157]

[0158] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0159] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for classifying orbital lymphoma and inflammatory pseudotumor, characterized in that, The classification method for orbital lymphoma and inflammatory pseudotumor includes: manually segmenting and preprocessing the tumor region and the ocular cone triangle region in the acquired orbital DCE-MRI images; dividing the sample dataset using the preprocessed 3D images as samples; extracting features from the tumor region using a neural network and performing cluster statistical analysis on the features as tumor region features; extracting features from the ocular cone triangle region as a similarity evaluation criterion, using a self-classification and self-recovery network to extract ocular cone region features; inputting the training dataset into the network and training the network model by combining the features of the tumor region and the ocular cone triangle region, and using the test set data for prediction; inputting multimodal data into the classification model for processing to achieve the classification task of orbital lymphoma and inflammatory pseudotumor. The classification method for orbital lymphoma and inflammatory pseudotumor includes the following steps: Step 1, Segmentation and Processing of Original Images: In the acquired orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original images are manually segmented using MITK software; the raw format data is processed into mat format data; the region of interest with the most values ​​greater than 0 is selected as the largest slice, and five slices are selected sequentially; the processed 3D original image, the 3D image of the tumor region, and the 3D image of the ocular cone triangle region are used as sample datasets, and training and testing sets are divided. Step 2, Feature extraction of the tumor region: Extract 9×9 image patches centered on each pixel of the segmented tumor region, and stack the extracted patches in the third dimension; build a two-layer neural network consisting of encoding and decoding parts, and train the neural network using the training dataset; input the training set and test set into the neural network to automatically extract features; use the k-means++ clustering algorithm to cluster the features into k classes; perform statistical analysis on the clustered features to obtain the tumor region features; Step 3, Feature extraction of the eye cone triangle region: Train an autoencoder network to extract deep features of the image as similarity to rank the data; generate five position coordinates and generate visual words at the five coordinate points; calculate the contrast loss function between the original image and the restored image through the self-recovery branch, and train a five-classification network; select the optimal model to obtain the deep feature extraction model of the eye cone triangle region, and extract the depth features of the eye cone region. Step 4, fusion feature classification based on ViT model: Input the training data of the original image into the pre-trained ViT model, combine the feature data obtained by the Transformer encoder with the texture features of the tumor and the depth features of the ocular cone triangle, and classify orbital lymphoma and inflammatory pseudotumor through the MLP classification head. Step 5, Multimodal Classification: The multimodal data is processed, and a soft voting strategy is used to integrate the modalities, ultimately completing the classification task for orbital lymphoma and inflammatory pseudotumor.

2. The classification method for orbital lymphoma and inflammatory pseudotumor as described in claim 1, characterized in that, Step one, in acquiring the orbital DCE-MRI image, includes segmentation and processing of the original image, which includes: (1) In the obtained orbital DCE-MRI images, the tumor region and the ocular cone triangle region in the original image were manually segmented using MITK software; (2) Convert the raw format data of the original image, the segmented tumor region and the eye cone triangle region into mat format data; (3) Count the number of values ​​greater than 0 in the image and select the region of interest with the most values ​​as the largest slice. Select two slices above and below the largest slice, for a total of five slices. (4) The processed three-dimensional original image, the three-dimensional image of the tumor region and the three-dimensional image of the eye cone triangle are used as sample datasets, and the training set and test set are randomly divided.

3. The method for classifying orbital lymphoma and inflammatory pseudotumor as described in claim 1, characterized in that, The feature extraction of the tumor region in step two includes: (1) Extract a 9×9 image patch block centered on each pixel of the segmented tumor region, and superimpose the extracted patch blocks in the third dimension; (2) Construct a two-layer neural network consisting of an encoding and a decoding part, and train the neural network using a training dataset; (3) Input the training set and test set into the trained neural network to automatically extract features; (4) Use the k-means++ clustering algorithm to cluster the features into k classes; (5) Statistical analysis of the clustered features yields the tumor region features.

4. The classification method for orbital lymphoma and inflammatory pseudotumor as described in claim 1, characterized in that, The feature extraction of the eye cone triangle region in step three includes: (1) Use the training set data to train the autoencoder network, and extract the deep features of the images as similarity to rank the data; (2) Generate five location coordinates and generate visual words at the five coordinate points; (3) Calculate the contrast loss function between the original image and the restored image through the self-recovery branch, and train the five-class classification network; wherein, the training is performed by minimizing the standard classification cross-entropy loss function, as shown in the following formula: ； In the formula, B represents the batch size; C represents the number of visual words; y and P represent the true value and the network prediction value, respectively; The recovery network is trained by minimizing the L2 distance between the original visual word and the reconstructed visual word, as shown in the following formula: ； In the formula, B represents the batch size, and x and x' represent the original visual word and the restored visual word, respectively; (4) Select the optimal model to obtain the deep feature extraction model of the eye cone triangle region; (5) Extract the depth features of the eye cone region.

5. The method for classifying orbital lymphoma and inflammatory pseudotumor as described in claim 1, characterized in that, Step four, which involves classification based on the ViT model, includes: (1) Input the training data of the original image into the pre-trained ViT model; (2) The feature data obtained after passing through the Transformer encoder is combined with the texture features of the tumor and the depth features of the ocular cone triangle; (3) Classification of orbital lymphoma and inflammatory pseudotumor using the MLP classification head; Step five involves processing the multimodal data to classify orbital lymphomas and inflammatory pseudotumors, including: (1) Perform data processing, tumor region feature extraction, eye cone region extraction and ViT network feature fusion steps on the three modal data T1WI, T1WI+C and T2WI respectively; (2) The results are classified into orbital lymphoma and inflammatory pseudotumor by using a soft voting multimodal integration strategy and the results are classified according to the probability statistics of the division.

6. An orbital lymphoma and inflammatory pseudotumor classification system applying the classification method for orbital lymphoma and inflammatory pseudotumor as described in any one of claims 1 to 5, characterized in that, The classification system for orbital lymphoma and inflammatory pseudotumor includes: The data acquisition module is used to acquire standard-compliant orbital DCE-MRI data; The data preprocessing module is used to manually segment the tumor region and the ocular cone triangle region in the acquired orbital DCE-MRI images using MITK software. The tumor region feature extraction module is used to extract a 9×9 image patch centered on each pixel of the segmented tumor region and automatically extract tumor region features to express tumor texture information. The eye cone triangle region feature extraction module is used to train the model with the eye cone triangle region image as input and to extract the depth features of the eye cone region using the eye cone triangle region deep feature extraction model. The ViT fusion feature classification module is used to input the training data of the original image into the pre-trained ViT model and fuse the feature data with the texture features of the tumor and the depth features of the ocular cone triangle. The multimodal classification module is used to process multimodal data, integrate modalities using a soft voting strategy, and complete the classification task of orbital lymphoma and inflammatory pseudotumor.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the classification method for orbital lymphoma and inflammatory pseudotumor as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the classification method for orbital lymphoma and inflammatory pseudotumor as described in any one of claims 1 to 5.

Images