Diagnostic system for intramedullary demyelinating pseudotumor disease and tumors based on predictive models

Abstract

The application discloses a kind of diagnosis system of intraspinal Demyelinating pseudotumor disease and tumor based on prediction model, belong to the field of construction of prediction model and system.A kind of diagnosis system of intraspinal Demyelinating pseudotumor disease and tumor, comprising: MRI scanning module, for obtaining the medical image of patient target area: image segmentation module, for obtaining the medical image in the MRI scanning module, and the medical image is preprocessed and segmented;Feature extraction module, for extracting features from the medical image processed in image segmentation module;Model prediction module, for receiving the extracted features, carries out prediction.

Description

Technical Field

[0001] This invention relates to the field of predictive model and system construction technology, specifically to a non-invasive diagnostic system for intramedullary demyelinating pseudotumor and tumors based on a predictive model. Background Technology

[0002] The existing techniques for differentiating intramedullary demyelinating lesions (TDLs) from intramedullary tumors mainly include the following three aspects, as well as their respective limitations:

[0003] 1. Imaging Differentiation: On non-contrast MRI, both intramedullary tumors and demyelinating pseudotumors may appear as high-signal areas on T2-weighted images, making them difficult to distinguish. On contrast-enhanced MRI, intramedullary tumors typically show enhancement, while the enhancement of demyelinating lesions may be subtle or exhibit an irregular pattern. Nevertheless, some demyelinating lesions may also show enhancement during the active phase, further complicating the differentiation. Current differential diagnosis methods include the use of characteristic images from traditional imaging techniques and novel sequence-assisted differentiation. Traditional characteristic image differentiation relies heavily on physician experience and lacks a systematic and reproducible diagnostic approach. While novel sequence-assisted differentiation may provide more information, it places high demands on medical equipment and is time-consuming, potentially limiting its application in clinical practice.

[0004] 2. Pathological Differentiation: This usually requires surgery or a biopsy to obtain a pathological specimen, which is then examined pathologically for diagnosis. However, this method is more invasive and carries a higher surgical risk. Furthermore, the results of the surgery or biopsy may take some time to arrive, potentially delaying the patient's treatment.

[0005] 3. Chemical indicator identification: This includes methods such as physical examination, blood tests, and cerebrospinal fluid analysis. These methods are typically time-consuming and expensive. Furthermore, they may require multiple sampling and testing, impacting the patient's quality of life.

[0006] In summary, based on the existing technologies described above, distinguishing between tumors and intramedullary demyelinating pseudotumors (TDLs) remains highly challenging. These issues pose a substantial difficulty in neurological diagnosis and treatment. In the absence of sufficient research, the accuracy and efficiency of diagnosing intramedullary TDL lesions are severely affected. This not only troubles physicians but also negatively impacts patient treatment and rehabilitation.

[0007] Therefore, existing technologies have some significant limitations and shortcomings in identification, and there is room for improvement. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a diagnostic system for intraspinal demyelinating pseudotumor and tumors based on a predictive model.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] A diagnostic system for intramedullary demyelinating pseudotumor and tumors, comprising:

[0011] The MRI scanning module is used to acquire medical images of the patient's target area.

[0012] An image segmentation module is used to acquire medical images from the MRI scanning module and to preprocess and segment the medical images.

[0013] The feature extraction module is used to extract features from the medical images processed by the image segmentation module;

[0014] The model prediction module is used to receive the extracted features and make predictions.

[0015] Optionally, it also includes a model interpretation module for predicting the contribution of features to the outcome prediction.

[0016] Optionally, the features include shape, size, position, texture, edges, and pixels.

[0017] Optionally, the features extracted by the feature extraction module include: first-order features, shape features, and texture features.

[0018] Optionally, the preprocessing includes medical image intensity normalization, format conversion, fixed resolution resampling, N4 magnetic field correction, and noise removal.

[0019] Optionally, the model prediction module is configured as an MLP composite model of T1WI and T2WI weighted images.

[0020] Another aspect of the present invention relates to a method for constructing a prediction model, comprising the following steps:

[0021] Feature filtering is performed, removing the feature with the highest redundancy in the current set during each iteration; multivariate selection is performed using minimum absolute shrinkage and selection operators;

[0022] Randomly allocate the data in the dataset to the training dataset and the test dataset;

[0023] Use different classifiers to build a single MR model or a combination of them.

[0024] Analyze and compare the prediction results of these models to select the optimal prediction model.

[0025] Optionally, the classifier includes LR, NaiveBayes, SVM, KNN, RF, ExtraTrees, XGBoost, LightGBM, GradientBoosting, and MLP.

[0026] A third aspect of the invention also relates to a prediction model constructed using the construction method described above.

[0027] The beneficial effects of this invention are:

[0028] 1. Innovative Feature Engineering: This invention not only incorporates traditional MRI image features such as shape, size, and location, but also integrates deeper radiomics features, including features encompassing texture, edge, and pixel meaning. These features greatly enrich the data and help improve the predictive performance of the model.

[0029] 2. Application of Ensemble Learning Models: This invention employs ensemble learning models, such as random forests or support vector machines, instead of single machine learning models. Ensemble models typically exhibit better generalization ability and prediction performance. By combining multiple weak learners, overfitting can be effectively reduced, and prediction accuracy can be improved.

[0030] 3. Enhanced Model Interpretability: By introducing interpretable machine learning techniques, such as SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Exlanations), this invention not only focuses on the model's predictive performance but also emphasizes its interpretability. This means we can understand the contribution of each feature to the model's predictions, which is crucial for clinical decision-making.

[0031] 4. Dynamic Model Update Mechanism: This invention employs a dynamic model update mechanism. As new patient data accumulates, the model can be updated and optimized periodically, ensuring that the model consistently maintains high predictive performance. Attached Figure Description

[0032] The invention will now be further described with reference to the accompanying drawings.

[0033] Figure 1 A flowchart illustrating the model construction process in the embodiments of this application;

[0034] Figure 2 To determine the types of features extracted and the sample distribution;

[0035] Figure 3 To use lasso regression for feature selection;

[0036] Figure 4 Construct a heatmap of AUC values ​​for the model group for the classifier;

[0037] Figure 5 The ROC curve and DCA curve of the MLP model;

[0038] Figure 6 The mixture matrix is ​​a comparison between the integrated MLP model and the radiologist's diagnosis;

[0039] Figure 7 To understand the contribution of 5 instances to model predictions using LIME. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] In an embodiment of the present invention, a method and application for constructing a predictive model for intramedullary demyelinating pseudotumor and tumors are disclosed, the general workflow of which is as follows: Figure 1 .

[0042] 1. Extraction of radiomics features

[0043] In this embodiment, medical imaging data of 75 patients with intramedullary tumors and TDL who visited Zhongda Hospital Affiliated to Southeast University between October 1, 2018 and August 31, 2023 were collected. The clinical data are shown in Table 1.

[0044] Table 1. Clinical data of 75 patients

[0045]

[0046] For continuous variables (age), t-tests were used for analysis, and for categorical variables (gender), Pearson's chi-square test was used. AST (Astrocytoma), EPN (Ependymoma), MS (Multiple Sclerosis), NMO (Neuromyelitis Optica) were also analyzed.

[0047] A set of radiomics features was extracted from MRI scans to characterize intraspinal demyelinating pseudotumors (TDLs) and tumors. This included the following steps:

[0048] 1) Convert the original image from DICOM format to NIfTI format, which is recognizable by the software;

[0049] 2) Due to different scanners or acquisition protocols, medical volumes often exhibit heterogeneous voxel spacing. This spacing relates to the physical interval between two pixels in an image. Spatial normalization is a common technique used to mitigate the impact of voxel spacing variability. In this embodiment, a fixed-resolution resampling method is employed (ensuring that images from different scanners or with different scanning parameters have a uniform spatial resolution. This ensures that extracted radiomics features are comparable and are not affected by the original image resolution) to address the aforementioned challenge. All images were resampled to normalize the voxel spacing to 333 mm.

[0050] 3) Image intensities may vary depending on the machine or scanning parameters. To avoid the impact of these differences on feature extraction, intensity normalization is required. This step uses the Z-score normalization method.

[0051] 4) Medical images often contain noise, such as noise caused by machine errors or biological motion. This noise may affect the feature extraction results, so Gaussian filtering is needed for noise removal.

[0052] 5) Use the N4 correction method for image correction.

[0053] 6) The target of image segmentation is the lesion region. When multiple lesions exist in the spinal cord, the largest lesion is selected as the volume of interest (VOI). On the PACS, axial T1WI (T1-weighted image) and T2WI (T2-weighted image) DICOM format images of each case are acquired. 3D Slicer software (HTTP: / / https: / / www.slicer.org; version 5.0.3) is used.

[0054] The volume of interest (VOI) was manually segmented. The intra-group correlation coefficient (ICC) for each feature was calculated to assess observer consistency and intra-observer reliability. Features exhibiting an ICC < 0.75 between and within observers were excluded.

[0055] 7) The extracted features can be divided into three different groups: (I) geometric features, (II) first-order features, and (III) texture features. The three-dimensional shape features of the tumor or lesion area are described by geometric attributes. The spatial distribution of intensity, inclusion patterns, and second- and higher-order features are described in detail by texture attributes. In this case, a variety of methods were used to extract texture attributes, including Gray Level Co-occurrence Matrix (GLCM) Features, Gray Level Size Zone Matrix (GLSZM) Features, Gray Level Run Length Matrix (GLRLM) Features, Neighbouring Gray Tone Difference Matrix (NGTDM) Features, and Gray Level Dependence Matrix (GLDM) Features. A total of 107 different categories of radiomics features were derived, including 18 geometric features, 14 first-order features, and 75 texture features. Pyradiomics was used.

[0056] The internal feature analysis program implemented at (http: / / pyradiomics.readthedocs.io) is used to extract all features. Feature grouping and distribution are as follows: Figure 2 .

[0057] 2. Construction and training of machine learning models

[0058] Using the aforementioned radiomics features, a machine learning model is constructed and trained. In this embodiment, the training data used during model training includes MRI data of TDLs with known diagnostic results and tumors. The goal of model training is to find a method that best utilizes radiomics features to distinguish between TDLs and tumors. This step aims to generate a model capable of accurately predicting TDLs and tumors. It includes the following steps:

[0059] 1) A t-test was performed for statistical analysis, and all radiomics features were screened. Only radiomics features with a p-value less than 0.05 were retained. Pearson rank correlation coefficients were used to calculate the correlation between features, and any feature with a correlation coefficient greater than 0.9 between any two attributes was retained.

[0060] 2) For feature filtering, a recursive elimination method is used, which requires removing the feature with the highest redundancy in the current set during each iteration. Subsequently, the Least Absolute Shrinkage and Selection Operator (LASSO) is used for multivariate selection, such as... Figure 3 We select the most suitable λ value and then choose a more valuable subset of features to build a classification model.

[0061] 3) The datasets are randomly distributed in an 8:2 ratio, allocated to either the training or testing datasets. The training dataset contains all cases used for model training, while the testing dataset is used to independently evaluate model performance.

[0062] 3. Model Validation and Optimization

[0063] The model performance is compared using multiple models and various sequence features. Validation is performed on independent test datasets to evaluate performance metrics such as accuracy, sensitivity, and specificity, leading to further adjustments and optimizations to the final model. This includes the following steps:

[0064] 1) In this embodiment, ten classifiers are used to construct a single MR sequence model (T1WI / T2WI) or a combined model (T1WI and T2WI).

[0065] 2) Subsequently, these models were compared. To evaluate the predictive performance of these 30 models, ROC curves were generated for the training and validation cohorts. The area under the ROC curve (AUC), accuracy, sensitivity, and specificity were calculated. Figure 4 The AUC values ​​of the 30 composite models constructed are shown.

[0066] 3) After comprehensive comparison, a composite MLP model combining T1WI and T2WI weighted images was selected. Figure 5 The ROC curve and DCA curve of the MLP model are shown.

[0067] 4) Radiologist A has 5 years of experience, and radiologist B has 3 years of experience. Neither radiologist is aware of the patient's diagnosis. Regarding information acquisition, both radiologists are aware of all the original images and other information input into the prediction model, and have the right to inquire about the patient's physical condition, symptoms, and other information. In other words, in the initial state, the radiologists have at least as much information as the prediction model constructed in this embodiment. For comparison, in this embodiment, an MLP composite model is selected for comparison with the radiologists' model. Table 2 shows the performance of the two radiologists and the radiomics model. Figure 6 The confusion matrix is ​​shown. The Delong test results indicate that the radiomics model significantly outperforms the radiologist model in both the training set and the full cohort (P<0.05).

[0068] Table 2. Diagnostic manifestations of radiologists and radiomics models

[0069]

[0070] 4. Pre-treatment diagnostic applications

[0071] Finally, the optimized model was applied to the differential diagnosis of intramedullary demyelinating pseudotumor and tumor before treatment. This step aims to translate the model into practical clinical application, helping physicians obtain more accurate diagnostic information before treatment and thus make better treatment decisions. It includes the following steps:

[0072] 1) First, new patient images need to be acquired from the medical device. These images are in DICOM format and include T1WI (T1-weighted images) and T2WI (T2-weighted images).

[0073] 2) Then, the same preprocessing steps are performed on the new patient images. These include image resampling, intensity normalization, noise removal, and N4 magnetic field correction.

[0074] 3) Next, features are extracted from the preprocessed images. These are exactly the same features used when training the model.

[0075] 4) Select features based on the training model results.

[0076] 5) Input the transformed features into the model to obtain the model's prediction result. This result can be a direct result or displayed using probability values, indicating the likelihood that the patient has intramedullary demyelinating pseudotumor or a tumor.

[0077] 6) Using LIME for model feature interpretation allows you to adjust feature weights based on the data's suitability. The LIME results for this example are as follows: Figure 7 .exist Figure 7 In LIME, a single prediction instance is selected, and a simulated dataset is generated around that instance, using five instances as examples. These simulated data points are weighted according to their similarity to the selected instance. LIME then trains a simple linear model (such as linear regression or logistic regression) to fit these simulated data. Finally, LIME interprets the weights of this simple model as a local interpretation of the original complex model. The weight of each feature represents the importance of that feature to the prediction result.

[0078] 7) The final diagnostic system typically requires combining the patient's clinical information with the physician's professional judgment. For example, if the model predicts an 80% probability that a patient has intraspinal tumors (TDL), but the patient's clinical symptoms do not match, the physician should choose further examination rather than making a direct diagnosis. Therefore, it is important to note that the diagnostic system described in this embodiment only provides a reference or intermediate result for distinguishing between intraspinal tumors and TDL; the final diagnostic result still requires the involvement of a professional.

[0079] These techniques are key components of this invention. They clearly describe how to extract useful radiomics features from MRI data, how to construct and train a machine learning model, and how to apply this model to actual clinical diagnosis. Through these techniques, this invention provides a novel, non-invasive, and systematic method for improving the accuracy of differential diagnosis between intramedullary demyelinating pseudotumors and tumors.

[0080] Compared to existing technologies, the diagnostic system of this invention combines radiomics and machine learning techniques, providing a novel, non-invasive, and systematic method to improve the accuracy of differential diagnosis between intramedullary demyelinating lesions (TDLs) and tumors. The main advantages and technical effects of this invention are as follows:

[0081] 1. Non-invasive: The method of this invention only requires MRI data and does not require surgery or biopsy, therefore it is non-invasive. This significantly reduces physical damage and surgical risks to patients.

[0082] 2. High Accuracy: By utilizing machine learning models, this invention can effectively leverage multiple radiomic features for disease identification. Compared to traditional imaging-based identification, this method offers significantly higher accuracy.

[0083] 3. Speed: Once the machine learning model is trained and optimized, it can identify new MRI data in a short time. This is faster than chemical indicator identification and pathological identification, providing patients with faster diagnostic results.

[0084] 4. Wide applicability: Since the method of the present invention only requires simple MRI data, it can be widely applied to various medical devices and cases, whether new or traditional machines.

[0085] (I) Experimental methods for prediction models

[0086] 1. Data Collection

[0087] First, a range of MRI data was collected, including data on intraspinal TDLs and tumors. This data was collected from multiple medical centers to increase data diversity. All MRI images underwent preprocessing, including denoising, standardization, and registration.

[0088] 2. Feature Extraction

[0089] For each MRI image, a series of radiomics features are extracted by delineating the VOI (Volume of Interest) of the lesion. These features include traditional features such as shape, size, and location, as well as deeper features such as texture, edges, and regions. These features can be extracted using existing image processing software or custom algorithms.

[0090] 3. Model Training and Validation

[0091] Then, we use the collected features and their corresponding disease labels (TDLs or tumors) to train our ensemble learning model. Cross-validation can be used to evaluate the model's performance and select the best model parameters.

[0092] 4. Model testing and optimization

[0093] Finally, we need to test the model's performance on a separate test set. If the performance is satisfactory, we can use the model for actual clinical diagnosis. Otherwise, we need to go back to the model training and validation steps to optimize the model.

[0094] 5. Data Interpretation

[0095] After the experiment, a series of data can be collected, including performance indicators such as model accuracy, sensitivity, and specificity. We can also collect interpretive data about the model, such as the LIME value of each feature. Analyzing this data allows us to understand the model's predictive power and the contribution of each feature in the prediction. This data can be used to support the technical solution of this invention, demonstrating its superiority in differentiating TDLs from tumors. Furthermore, visually demonstrating the contribution of each feature can also provide a reference for doctors' diagnoses.

[0096] 6. Pre-treatment diagnostic applications

[0097] Finally, we applied the optimized model to the differential diagnosis of intramedullary demyelinating pseudotumor and intramedullary tumors before treatment. This step aims to translate our invention into practical clinical application, helping physicians obtain more accurate diagnostic information before treatment and thus make better treatment decisions.

[0098] (II) Implementation Examples

[0099] 1. Data preparation and preprocessing

[0100] 1) Medical imaging data of 75 patients with intramedullary tumors and TDL who visited Zhongda Hospital Affiliated to Southeast University between October 1, 2018 and August 31, 2023 have been collected.

[0101] 2) Convert the DICOM image format of these patients' MRI scans to NIfTI format.

[0102] 3) Use a fixed-resolution resampling method for spatial normalization to mitigate the impact of voxel spacing heterogeneity.

[0103] 4) Normalize the intensity of MRI images to eliminate intensity differences caused by different scanning parameters or the machine.

[0104] 5) Perform noise removal on MRI images to eliminate noise that may affect feature extraction.

[0105] 6) Perform N4 correction on MRI images.

[0106] 7) Segment the volume of interest (VOI), especially when there are multiple lesions, select the largest lesion as the VOI.

[0107] 2. Radiomics Feature Extraction

[0108] 1) Use the pyradiomics package to extract radiomics features, including geometric features, first-order features, and texture features.

[0109] 3. Feature selection and machine learning model construction

[0110] 1) Perform a T-test statistical check and delete features with p-values ​​greater than 0.05.

[0111] 2) Calculate the correlation between features and retain any feature whose correlation coefficient between any two attributes exceeds 0.9.

[0112] 3) Use recursive elimination for feature filtering, and then use LASSO for multivariate selection.

[0113] 4) Divide the dataset, for example, by allocating it to the training dataset and the test dataset in an 8:2 ratio.

[0114] 5) Build and train machine learning models based on the selected features.

[0115] 4. Model Validation and Optimization

[0116] 1) Use ten classifiers (LR, Logistic Regression; NaiveBayes, Naive BayesClassifier; SVM, Support Vector Machine; KNN:K Nearest Neighbors Classifier; RF, Random Forest Classifier; ExtraTrees, Extra Trees Classifier; Classifier; MLP, Multi-LayerPerceptron Classifier.) built the model.

[0117] 2) Compare the predictive performance of different models, and calculate AUC, accuracy, sensitivity and specificity by generating ROC curves.

[0118] 3) Select the optimal model based on the comparison results. For this dataset, an MLP model using composite T1 & T2WI sequences was selected.

[0119] 5. Explanation of Model Features

[0120] 1) Use LIME (Locally Interpretable Model Interpretation) to interpret model predictions. LIME provides a set of weights for each prediction, illustrating the contribution of each feature to the prediction result.

[0121] 2) By analyzing the feature weights provided by LIME, we can further understand the working mechanism and decision-making process of the model.

[0122] The significance of using LIME is that it provides transparency to the model, helping us understand the model's predictive behavior.

[0123] 6. Auxiliary diagnosis

[0124] Doctors can combine their professional knowledge with the model's predictions to make a final diagnosis. The advantage of this machine learning model lies in:

[0125] Non-invasive: While there is a gold standard for diagnosis in clinical practice, these standards often require invasive methods, such as biopsies or slides, to obtain pathological results. This can be a significant obstacle for patients with mild symptoms or those concerned about surgical risks. Our model, however, can make predictions using only MRI images, greatly reducing the physical burden on patients.

[0126] High efficiency: The predictive model of this invention can complete the prediction within minutes, which greatly improves the efficiency of diagnosis.

[0127] Early warning: Before the gold standard diagnostic results are available, our model can provide early warning information to help doctors develop treatment plans in advance and improve the quality of treatment for patients.

[0128] Explanatory power: The predictive model of this invention can not only make predictions, but also explain the role of each feature in the prediction, providing more diagnostic information and helping doctors understand the complexity of the condition.

[0129] Accuracy: Our model has been trained on a large amount of data, and its prediction accuracy may be comparable to the gold standard diagnosis and not inferior to that of young radiologists.

[0130] Humanistic Care: By providing faster predictive results, our model can reduce patients' anxiety and uncertainty while waiting for diagnostic results. Furthermore, early warnings and personalized predictive results can make patients feel more cared for and understood, improving their healthcare experience and thus achieving the goal of humanistic care.

[0131] Therefore, despite the existence of a diagnostic gold standard, the machine learning model of this invention can still play an important auxiliary role in clinical diagnosis, improving diagnostic efficiency and accuracy, providing more diagnostic information, reducing the physical burden on patients, and providing better patient care.

[0132] The diagnostic prediction model provided in this embodiment has demonstrated excellent diagnostic and treatment effects in practice. More importantly, the method in this embodiment is significantly superior to the results obtained by junior radiologists in assisting diagnosis. This means that by applying the prediction model in this embodiment, the accuracy and efficiency of diagnosis can be improved, alleviating the difficulties in neurological diagnosis and treatment, and ultimately enhancing patients' treatment outcomes and quality of life.

[0133] In summary, the technical solution of this invention fills a research gap in the field, resolves existing diagnostic and treatment difficulties, and demonstrates significant advantages in practice. We expect this research to further promote the development of this field and bring substantial improvements to neurological diagnosis and treatment.

[0134] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0135] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A diagnostic system for intramedullary demyelinating pseudotumor and tumors, characterized in that, include: MRI scanning module, Used to acquire magnetic resonance medical images of the spinal cord region of a patient, the medical images including T1WI weighted imaging sequences and T2WI weighted imaging sequences under plain magnetic resonance imaging conditions; The image segmentation module is used to receive medical images acquired by the MRI scanning module, perform preprocessing and segmentation on the medical images to obtain the volume of interest for analysis of intraspinal lesions; The preprocessing includes image intensity normalization, format conversion, fixed resolution resampling, N4 magnetic field inhomogeneity correction, and noise suppression. When multiple lesion regions exist in the spinal cord, the lesion region with the largest volume is selected as the volume of interest. The segmentation includes segmenting the volume of interest and evaluating inter-observer consistency and intra-observer reliability by calculating the intra-group correlation coefficient of features, thereby filtering out features with insufficient consistency. The feature extraction module is used to extract image features characterizing the properties of intraspinal lesions from the volume of interest processed by the image segmentation module. The features extracted by the feature extraction module include: first-order features, shape features, and texture features; texture features include gray-level co-occurrence matrix features, gray-level size region matrix features, gray-level run-length matrix features, adjacent gray-level difference matrix features, and gray-level dependency matrix features. The model prediction module is used to receive the image features output by the feature extraction module, perform modeling and analysis of intraspinal lesions, and output the differentiation results between intraspinal demyelinating pseudotumor and tumor. The model prediction module is configured as a multilayer perceptron composite model based on weighted image features combining T1WI and T2WI.

2. The diagnostic system for intramedullary demyelinating pseudotumor and tumors according to claim 1, characterized in that, It also includes a model interpretation module, which is used to predict the contribution of features to the result prediction.

3. The diagnostic system for intramedullary demyelinating pseudotumor and tumors according to claim 1, characterized in that, The features include shape, size, position, texture, edges, and regions.

Images