Method and system for evaluating thyroid eye disease activity based on MRI (Magnetic Resonance Imaging)

Abstract

The invention discloses an MRI (Magnetic Resonance Imaging)-based thyroid eye disease activity assessment method and system, and aims to solve the problems that the deep state is difficult to assess and the subjectivity is strong in the existing clinical activity score. The method comprises the following steps: acquiring and preprocessing a magnetic resonance image and generating a standardized three-dimensional image; automatically segmenting an orbit structure through a deep learning network; extracting multi-parameter image features; calculating a disease activity probability score by using a machine learning classifier; and mapping the probability score into an activity grading result and generating an evaluation report. The system comprises an image data acquisition and preprocessing module, an orbit structure automatic segmentation module, a multi-parameter image feature extraction module, an activity evaluation model construction and reasoning module and a clinical report generation module. According to the technical scheme, image evaluation efficiency and repeatability can be remarkably improved, subjective deviation is reduced, active-stage inflammatory edema and inactive-stage fibrosis are accurately and quantitatively distinguished, evaluation specificity is improved, and support is provided for individualized treatment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a thyroid eye disease activity evaluation method and system based on MRI. BACKGROUND

[0002] Thyroid eye disease is a disabling orbital disease closely related to autoimmunity, and its clinical course is usually divided into active and inactive stages. There are essential differences in pathological characteristics, treatment response and intervention strategies between different stages. The core pathological mechanism of the disease involves inflammatory infiltration, edema and subsequent fibrosis of orbital fat tissue and extraocular muscles. Accurate determination of the stage of the disease is of decisive significance for developing individualized treatment plans. Currently, the disease activity is mainly evaluated by clinical activity score in clinical practice. Although this method is simple to operate, its evaluation dimension is limited and it is difficult to fully reflect the continuity of disease dynamic evolution.

[0003] Among them, non-invasive evaluation technology based on magnetic resonance imaging is considered as a potential alternative or supplementary method due to its good soft tissue resolution and repeatability. Magnetic resonance imaging can directly show key pathological signs such as extraocular muscle thickening, signal abnormalities and fat tissue changes, and theoretically has the potential to reflect the degree of inflammation activity. However, existing image evaluation methods are mostly qualitative or semi-quantitative, and lack a quantitative analysis framework specifically constructed for the activity of thyroid eye disease, making it difficult to accurately quantify the inflammatory load, edema degree and tissue microstructure changes.

[0004] The existing technology still has significant deficiencies in evaluating the activity of thyroid eye disease using magnetic resonance imaging: first, there is a lack of automatic segmentation and feature extraction mechanism for specific anatomical structures of the orbit, resulting in a high dependence on manual intervention in the analysis process, low efficiency and poor repeatability; second, there is no multi-parameter image biomarker system directly related to disease activity, which cannot effectively distinguish between inflammatory edema in the active stage and fibrosis changes in the inactive stage; third, existing models generally ignore the epidemiology and image phenotype characteristics of the Chinese population, and directly applying western standards may introduce systematic bias; finally, there is no end-to-end intelligent evaluation system, making it difficult to achieve fast, objective and interpretable activity classification in clinical routine examination. The above problems seriously restrict the application value of magnetic resonance imaging in the dynamic monitoring and precise diagnosis and treatment of thyroid eye disease, and an intelligent evaluation method and system that can deeply integrate image features and clinical needs is urgently needed. SUMMARY

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an MRI-based method and system for assessing the activity of thyroid eye disease, which can effectively solve the problems in the background technology. To achieve the above objective, this invention provides the following technical solution: On one hand, an MRI-based system for assessing the activity of thyroid eye disease, comprising the following components: an image data acquisition and preprocessing module, used to acquire and standardize orbital magnetic resonance imaging sequences of patients with thyroid eye disease, generating standardized three-dimensional image data; an automated orbital structure segmentation module, connected to the image data acquisition and preprocessing module, used to receive the standardized three-dimensional image data and perform precise pixel-level segmentation of the extraocular muscles, optic nerve, and adipose tissue within the orbit based on a deep learning segmentation network, generating a segmentation mask; and a multi-parameter image feature extraction module, connected to the automated orbital structure segmentation module... An automated segmentation module is used to quantitatively extract morphological features, signal intensity features, and texture features related to thyroid eye disease activity from the standardized 3D image data based on the segmentation mask, forming a high-dimensional feature vector. An activity assessment model construction and inference module, connected to the multi-parameter image feature extraction module, is used to receive the high-dimensional feature vector and calculate a probability score representing disease activity based on a pre-trained machine learning classifier. A clinical decision support and report generation module, connected to the activity assessment model construction and inference module, is used to map the probability score to a clinically interpretable activity grading result and automatically generate a structured assessment report.

[0006] Preferably, the image data acquisition and preprocessing module specifically performs the following operations: receiving raw image data from a DICOM-compliant magnetic resonance imaging (MRI) device; performing intensity normalization processing on the raw image data to map the image signal intensity under different scanning devices and protocols to a standardized grayscale range; performing three-dimensional spatial registration to align all sequences to the same anatomical coordinate system to eliminate spatial deviations caused by patient position and movement; and performing isotropic resampling to unify the voxel size to 1mm×1mm×1mm to ensure scale consistency in subsequent analysis.

[0007] Furthermore, the automated orbital structure segmentation module employs a deep learning segmentation network based on an improved 3DU-Net architecture. This network embeds residual connections in the encoding path to mitigate gradient vanishing and introduces an attention gating mechanism in the decoding path to focus on key orbital anatomical structures. The training data for the segmentation network comes from a dataset of orbital structures of thyroid eye diseases, manually annotated by professional radiologists and containing imaging characteristics of the Chinese population. The segmentation mask output by the segmentation network undergoes morphological post-processing to eliminate discontinuities in segmentation boundaries and internal holes, ensuring the topological correctness of the segmentation results.

[0008] Furthermore, the morphological features extracted by the multi-parameter image feature extraction module include the maximum cross-sectional area, volume, diameter ratio, and curvature features of the extraocular muscles; signal intensity features include the signal intensity ratio of the extraocular muscles to the ipsilateral temporalis muscle on T2-weighted images, and the signal suppression rate of orbital fat tissue on fat suppression sequences; texture features are calculated using the gray-level co-occurrence matrix and run length matrix, covering indicators such as contrast, correlation, energy, homogeneity, and short run advantage; before being input into the activity assessment model, the high-dimensional feature vectors need to undergo feature selection based on the maximum correlation and minimum redundancy criterion to reduce dimensionality and improve the model's generalization ability.

[0009] Preferably, the pre-trained machine learning classifier in the activity assessment model construction and inference module is a support vector machine, and its kernel function is a radial basis function. The training process of the support vector machine adopts grid search and 5-fold cross-validation for hyperparameter optimization. The hyperparameters include the penalty coefficient C and the kernel function parameter γ. The probability score is calculated based on the decision function value output by the support vector machine, which is calibrated by the sigmoid function. Its value range is 0 to 1, and the higher the value, the greater the probability that the disease is in an active phase.

[0010] Furthermore, the specific rules for mapping probability scores to activity grading results in the clinical decision support and report generation module are as follows: when the probability score is below 0.3, it is determined to be in the clinical inactive period; when the probability score is between 0.3 and 0.7, it is determined to be in the borderline state; when the probability score is above 0.7, it is determined to be in the clinical active period; the structured assessment report automatically integrates key imaging feature values, activity grading results, and comparative analysis charts with previous examinations.

[0011] On the other hand, an MRI-based method for assessing the activity of thyroid eye disease includes the following steps: Step S110, acquiring raw orbital magnetic resonance imaging data of patients with thyroid eye disease, and performing intensity normalization, spatial registration, and isotropic resampling on the raw data to generate standardized three-dimensional image data; Step S120, inputting the standardized three-dimensional image data into a pre-trained deep learning segmentation network to automatically segment the extraocular muscles, optic nerve, and adipose tissue within the orbit, and outputting the corresponding segmentation mask; Step S130, based on the segmentation mask, quantitatively extracting a set of predefined multi-parameter image features from the standardized three-dimensional image data, the feature set covering morphological, signal intensity, and texture feature dimensions; Step S140, inputting the feature vector composed of the multi-parameter image features into a pre-trained activity assessment classification model to calculate a probability score characterizing the activity of thyroid eye disease; Step S150, converting the probability score into a specific activity level according to a preset threshold rule, and automatically generating a clinical assessment report containing quantitative features and level conclusions.

[0012] Compared with existing technologies, this invention has the following beneficial effects: 1. Through an integrated and automated segmentation and feature extraction process, it significantly improves the analytical efficiency and repeatability of thyroid eye disease image assessment, reducing subjective bias introduced by manual intervention; 2. It constructs a multi-parameter imaging biomarker system closely related to the pathophysiological mechanism of the disease, achieving accurate quantitative differentiation between active inflammatory edema and inactive fibrotic changes, thus improving diagnostic specificity; 3. Based on the imaging phenotypic characteristics of the Chinese population, the model is trained and optimized, effectively avoiding the systematic bias that may be introduced by directly applying Western standards, and improving the applicability and accuracy of the model in Chinese clinical practice; 4. It forms an end-to-end intelligent assessment closed loop, which can provide rapid, objective and clinically interpretable activity grading in routine clinical examination scenarios, strongly supporting individualized treatment decisions for thyroid eye diseases. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall technical architecture of a method and system for assessing the activity of thyroid eye disease based on MRI, as proposed in this invention.

[0014] Figure 2 This is a schematic diagram of the core principle framework of the automated orbital structure segmentation based on the 3D U-Net architecture in this invention;

[0015] Figure 3 This is a logical flowchart of the multi-parameter image feature extraction module in this invention;

[0016] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the activity assessment model construction and reasoning module in this invention;

[0017] Figure 5 This is a comparison diagram of the technical effects / principles of the clinical decision support and report generation module in this invention; Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments according to the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0019] Example 1

[0020] In the endocrinology outpatient department of a tertiary-level Class A hospital, an MRI-based thyroid eye disease activity assessment system is used to conduct comprehensive imaging evaluations for patients diagnosed with thyroid eye disease. (See also...) Figure 1This system includes an image data acquisition and preprocessing module, an automated orbital structure segmentation module, a multi-parameter image feature extraction module, an activity assessment model construction and inference module, and a clinical decision support and report generation module. The image data acquisition and preprocessing module receives raw image data from DICOM-compliant MRI equipment, including various scanning sequences such as T1-weighted sequences, T2-weighted sequences, and fat-suppressed sequences. The raw image data undergoes intensity normalization, mapping the image signal intensity from different scanning equipment and protocols to a standardized grayscale range of 0 to 255, eliminating signal differences between equipment. Three-dimensional spatial registration is performed, aligning all sequences to the same anatomical coordinate system to eliminate spatial deviations caused by patient position and movement; registration accuracy is controlled to the sub-millimeter level. Isotropic resampling is performed, unifying voxel sizes to 1 mm × 1 mm × 1 mm to ensure scale consistency in subsequent analysis; the resampling process uses a cubic spline interpolation algorithm to maintain image quality.

[0021] The automated orbital structure segmentation module receives standardized 3D image data and performs precise pixel-level segmentation of the extraocular muscles, optic nerve, and adipose tissue within the orbit based on a deep learning segmentation network. See also... Figure 2 This deep learning segmentation network employs an improved model based on the 3D U-Net architecture. Residual connections are embedded in the encoding path to alleviate the vanishing gradient problem, and an attention gating mechanism is introduced in the decoding path to focus on key orbital anatomical structures. The training data for the segmentation network comes from a dataset of orbital structures in thyroid eye diseases, manually annotated by professional radiologists and containing image characteristics of the Chinese population. This dataset covers 1500 image sequences from 500 patients. The segmentation mask output by the segmentation network undergoes morphological post-processing, employing a combination of opening and closing operations to eliminate discontinuities and internal holes at segmentation boundaries, ensuring the topological correctness of the segmentation results. The segmentation accuracy achieves a Descein similarity coefficient of over 0.92.

[0022] The multi-parameter image feature extraction module, based on segmentation masking, quantitatively extracts multi-dimensional features related to thyroid eye disease activity from standardized 3D image data. (See also...) Figure 3Morphological feature extraction includes calculating the maximum cross-sectional area of ​​the extraocular muscles by measuring the maximum cross-sectional contour of each extraocular muscle on axial images; volume measurement by accumulating the actual spatial volume of all voxels within the segmented region; diameter ratio calculation including the ratio of the transverse to longitudinal diameter of the extraocular muscles and the ratio of muscle thickness to width; curvature features obtained by calculating the histogram of curvature distribution on the surface of the extraocular muscles. Signal intensity feature extraction includes calculating the signal intensity ratio of the extraocular muscles to the ipsilateral temporalis muscle on T2-weighted images, selecting the region with the highest signal intensity among the extraocular muscles for comparison with the normal region of the ipsilateral temporalis muscle; and calculating the signal inhibition rate of orbital fat tissue on fat-suppressed sequences by comparing the percentage change in signal intensity before and after fat suppression. Texture features are calculated using a gray-level co-occurrence matrix and a run-length matrix. The gray-level co-occurrence matrix calculates four features in 13 directions: contrast, correlation, energy, and homogeneity. The run-length matrix calculates five features: short run advantage, long run advantage, and gray-level non-uniformity. All extracted features constitute a high-dimensional feature vector with a feature dimension of 189. Before being input into the activity assessment model, feature selection based on the maximum relevance and minimum redundancy criterion is required. By calculating the mutual information between each feature and the activity label, as well as the redundancy between features, the top 30 most discriminative features are retained.

[0023] The activity assessment model building and inference module receives a high-dimensional feature vector after feature selection and calculates a probability score representing disease activity based on a pre-trained machine learning classifier. See also Figure 4 This module employs a Support Vector Machine (SVM) as its machine learning classifier, with a radial basis function (RBF) as its kernel function. The SVM training process utilizes grid search and 5-fold cross-validation for hyperparameter optimization. The hyperparameters include the penalty coefficient C and the kernel parameter γ, with C ranging from 0.1 to 100 in the logarithmic space and γ from 0.001 to 1 in the logarithmic space, respectively. The probability score is calculated based on the decision function value output by the SVM, calibrated using the Sigmoid function. Its value ranges from 0 to 1, with higher values ​​indicating a greater likelihood of the disease being in an active phase. The dataset used for model training includes 300 clinically diagnosed patients with thyroid ophthalmopathy, 150 in the active phase and 150 in the inactive phase. The model achieved an area under the receiver operating characteristic (AUC) of 0.89 on the independent test set.

[0024] The clinical decision support and report generation module maps probability scores to clinically interpretable activity grading results. The specific mapping rules are as follows: a probability score below 0.3 is considered clinically inactive; a probability score between 0.3 and 0.7 is considered borderline; and a probability score above 0.7 is considered clinically active. See also... Figure 5The structured assessment report automatically integrates key imaging feature values, activity grading results, and comparative analysis charts with previous examinations. The report generation process first extracts key parameters from the current examination, including extraocular muscle volume, T2 signal intensity ratio, and texture feature values; then it compares these parameters with the patient's historical examination data over time to calculate the trends of each parameter; finally, it generates a comprehensive assessment report containing text descriptions, data tables, and multi-parameter trend charts, with a report format conforming to the hospital's electronic medical record system integration standards.

[0025] In the specific clinical workflow, step S110 performs image data acquisition and preprocessing operations, receiving raw DICOM data from the MRI equipment, performing intensity normalization processing, with the normalization parameters dynamically adjusted according to the equipment model and scanning protocol; performing spatial registration operations, using the midline structure as a reference for six-degree-of-freedom rigid body transformation; and performing isotropic resampling, with the resampling algorithm ensuring no loss of image details. Step S120 performs automated orbital structure segmentation, inputting standardized 3D image data into a pre-trained deep learning segmentation network, with network inference time controlled within 2 minutes; outputting segmentation masks for extraocular muscles, optic nerve, and adipose tissue, and verifying the segmentation results through 3D visualization. Step S130 performs multi-parameter image feature extraction, locating each anatomical structure region based on the segmentation mask; calculating morphological features, signal intensity features, and texture features respectively; the feature extraction process uses a parallel computing architecture to improve efficiency. Step S140 performs activity assessment calculation, inputting feature vectors into an activity assessment classification model; model inference is based on a support vector machine decision function; outputting probability scores and recording confidence intervals. Step S150 generates a clinical report, determines the activity level based on the probability score, integrates all quantitative features to generate a structured report, and automatically transmits the report to the attending physician's workstation through the hospital information system.

[0026] Example 2

[0027] In the thyroid department of a regional medical center, an MRI-based thyroid ophthalmopathy activity assessment system provides personalized assessment services for patients at different stages of the disease. The image data acquisition and preprocessing module supports data access from various MRI equipment models, including 1.5T and 3.0T devices from major manufacturers such as Siemens, GE, and Philips. Intensity normalization processing employs an adaptive normalization algorithm based on tissue signals, establishing a device-independent standardized grayscale mapping relationship by identifying the signal intensity of reference tissues such as cerebrospinal fluid and muscle in the image. Spatial registration uses a multimodal registration method based on mutual information, achieving a registration accuracy within 0.5 mm, ensuring the comparability of examination results at different time points.

[0028] The automated orbital structure segmentation module, based on the fundamental 3D U-Net architecture, adds a multi-scale feature fusion mechanism. Features extracted at different levels of the encoder are skipped in the decoder, enhancing the ability to recognize small anatomical structures. The segmentation network is trained using a transfer learning strategy, first pre-trained on a large natural image dataset, and then fine-tuned on a thyroid eye disease-specific dataset to improve the model's generalization performance with limited labeled data. The segmentation results undergo post-processing based on connected component analysis to automatically identify and repair discrete outliers in the segmented regions, ensuring the continuity of the segmentation mask.

[0029] The multi-parameter image feature extraction module, building upon the basic feature set, adds time-signal intensity curve feature analysis for dynamic contrast-enhanced scanning. For patients undergoing dynamic contrast-enhanced scanning, it extracts the signal intensity change curve of the extraocular muscles after contrast agent injection, calculating dynamic parameters such as area under the curve, peak time, and maximum enhancement rate. Texture feature analysis is extended to three-dimensional space, obtaining spatial texture features by calculating the three-dimensional gray-level co-occurrence matrix, providing a more comprehensive characterization of tissue microstructure changes. The feature selection process employs a recursive feature elimination algorithm, iteratively training the model and removing the least important features to ultimately determine the optimal feature subset.

[0030] The activity assessment model building and inference module supports the integrated evaluation of multiple machine learning algorithms. In addition to Support Vector Machines, it also integrates Random Forest, Gradient Boosting Decision Tree, and other algorithms. A voting mechanism is used to synthesize the prediction results of each algorithm, improving the robustness of the evaluation. The model update mechanism supports incremental learning; when new labeled data is added, incremental training can be performed on the existing model, avoiding the time cost of retraining. The probability score calculation introduces uncertainty estimation, providing a confidence interval for the prediction results by calculating the model's dropout variance during testing, thus assisting clinical decision-making.

[0031] The clinical decision support and report generation module has added a personalized treatment recommendation generation function, automatically generating targeted treatment recommendations based on the activity grading results and specific characteristics. For patients in the active phase, glucocorticoid pulse therapy is recommended with a specified dosage reference range; for patients in the borderline phase, close follow-up is recommended with the follow-up interval specified; for patients in the inactive phase, rehabilitation therapy or surgical correction is recommended. The report generation supports multilingual output to meet the reading needs of patients in different regions, and also provides a simplified patient version of the report, explaining the test results and clinical significance in plain language.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method and system for assessing the activity of thyroid ophthalmopathy based on MRI, characterized in that, The system comprises the following components: an image data acquisition and preprocessing module, used to receive raw image data from a DICOM-compliant magnetic resonance imaging (MRI) device, perform intensity normalization on the raw image data, map the image signal intensity under different scanning devices and protocols to a standardized grayscale range, perform three-dimensional spatial registration to align all sequences to the same anatomical coordinate system to eliminate spatial deviations caused by patient position and movement, perform isotropic resampling to unify voxel sizes to 1 mm × 1 mm × 1 mm, and generate standardized three-dimensional image data; and an automated orbital structure segmentation module, connected to the image data acquisition and preprocessing module, used to receive the standardized three-dimensional image data and perform precise pixel-level segmentation of the extraocular muscles, optic nerve, and adipose tissue within the orbit based on a deep learning segmentation network, wherein the deep learning segmentation network is based on 3D... An improved model based on the U-Net architecture embeds residual connections in the encoding path to mitigate gradient vanishing and introduces an attention gating mechanism in the decoding path to focus on key orbital anatomical structures. The segmentation mask output by the segmentation network undergoes morphological post-processing to eliminate discontinuities and internal holes at segmentation boundaries, generating a segmentation mask. A multi-parameter image feature extraction module, connected to the automated orbital structure segmentation module, is used to quantitatively extract morphological features, signal intensity features, and texture features related to thyroid eye disease activity from the standardized 3D image data based on the segmentation mask. The morphological features include the maximum cross-sectional area, volume, diameter ratio, and curvature features of the extraocular muscles. The signal intensity features include the signal intensity ratio of the extraocular muscles to the ipsilateral temporalis muscle on T2-weighted images and the signal suppression rate of orbital fat tissue on fat suppression sequences. The texture features are calculated using a gray-level co-occurrence matrix and a run-length matrix. The system comprises a high-dimensional feature vector, encompassing indicators such as contrast, correlation, energy, homogeneity, and short run advantage. An activity assessment model construction and inference module, connected to a multi-parameter image feature extraction module, receives the high-dimensional feature vector and calculates a probability score representing disease activity based on a pre-trained machine learning classifier. The machine learning classifier is a support vector machine (SVM) with a radial basis function as its kernel. The SVM training process employs grid search and 5-fold cross-validation for hyperparameter optimization, with hyperparameters including a penalty coefficient C and a kernel function parameter γ. The probability score is calculated based on the decision function value output by the SVM, calibrated using a sigmoid function. A clinical decision support and report generation module, connected to the activity assessment model construction and inference module, maps the probability score to clinically interpretable activity grading results and automatically generates a structured assessment report.

2. The MRI-based thyroid ophthalmopathy activity assessment system according to claim 1, characterized in that, In the image data acquisition and preprocessing module, intensity normalization mapping maps the image signal intensity to a standardized grayscale range of 0 to 255. The registration accuracy of three-dimensional spatial registration is controlled at the sub-millimeter level. The isotropic resampling process uses a cubic spline interpolation algorithm to maintain image quality.

3. The MRI-based thyroid ophthalmopathy activity assessment system according to claim 1, characterized in that, In the automated orbital structure segmentation module, the training data of the deep learning segmentation network comes from a dataset of orbital structures of thyroid eye diseases that are manually annotated by professional radiologists and contain image features of the Chinese population. The morphological post-processing of the segmentation mask adopts a combination of opening and closing operations, and the segmentation accuracy reaches a Dessian similarity coefficient of over 0.

92.

4. The MRI-based thyroid ophthalmopathy activity assessment system according to claim 1, characterized in that, In the multi-parameter image feature extraction module, the high-dimensional feature vector needs to undergo feature selection based on the maximum correlation and minimum redundancy criterion before being input into the activity assessment model. By calculating the mutual information between each feature and the activity label, as well as the redundancy between features, the top 30 most discriminative features are retained.

5. The MRI-based thyroid ophthalmopathy activity assessment system according to claim 1, characterized in that, In the activity assessment model construction and inference module, the hyperparameter search range of the support vector machine is the logarithmic space of C from 0.1 to 100 and the logarithmic space of γ from 0.001 to 1. The dataset used for model training contains 300 clinically diagnosed patients with thyroid eye disease, of which 150 are in the active phase and 150 are in the inactive phase.

6. The MRI-based thyroid ophthalmopathy activity assessment system according to claim 1, characterized in that, The specific rules for mapping probability scores to activity grading results in the clinical decision support and report generation module are as follows: when the probability score is below 0.3, it is determined to be in the clinical inactive period; when the probability score is between 0.3 and 0.7, it is determined to be in the borderline state; when the probability score is above 0.7, it is determined to be in the clinical active period.

7. A method for assessing the activity of thyroid ophthalmopathy based on MRI, characterized in that, The method includes the following steps: Step S110, acquiring raw orbital magnetic resonance imaging data of patients with thyroid eye disease, and performing intensity normalization, spatial registration, and isotropic resampling on the raw data to generate standardized three-dimensional image data; Step S120, inputting the standardized three-dimensional image data into a pre-trained deep learning segmentation network to automatically segment the extraocular muscles, optic nerve, and adipose tissue within the orbit, and outputting the corresponding segmentation mask; Step S130, based on the segmentation mask, quantitatively extracting a set of predefined multi-parameter image features from the standardized three-dimensional image data, the feature set covering morphological features, signal intensity features, and texture features; Step S140, inputting the feature vector composed of the multi-parameter image features into a pre-trained activity assessment classification model to calculate a probability score characterizing the activity of thyroid eye disease; Step S150, converting the probability score into a specific activity level according to a preset threshold rule, and automatically generating a clinical assessment report containing quantitative features and level conclusions.

8. The MRI-based method for assessing thyroid eye disease activity according to claim 7, characterized in that, In step S110, intensity normalization maps the image signal intensity to a standardized grayscale range of 0 to 255. Spatial registration uses the midline structure as a reference for a six-degree-of-freedom rigid body transformation. Isotropic resampling uses a cubic spline interpolation algorithm.

9. The MRI-based method for assessing thyroid eye disease activity according to claim 7, characterized in that, In step S130, the morphological features include the maximum cross-sectional area, volume, diameter ratio, and curvature features of the extraocular muscles; the signal intensity features include the signal intensity ratio of the extraocular muscles to the ipsilateral temporalis muscle on T2-weighted images and the signal inhibition rate of orbital fat tissue on fat inhibition sequences; and the texture features are calculated using the gray-level co-occurrence matrix and the run length matrix.

10. The MRI-based method for assessing the activity of thyroid ophthalmopathy according to claim 7, characterized in that, In step S140, the activity assessment classification model is a support vector machine, and its kernel function is a radial basis function. The probability score is calculated based on the decision function value output by the support vector machine, which is calibrated by the sigmoid function, and its value range is 0 to 1.

Images