Thyroid eye disease activity evaluation method and system based on eye image

Abstract

The invention discloses a thyroid eye disease activity evaluation method and system based on an eye image, and aims to solve the problems that the existing clinical activity score cannot reflect continuous change of an illness state and is susceptible to subjective influence. The method comprises the following steps: acquiring original image data containing an eye region; performing standardized preprocessing on the image data; extracting multi-scale inflammatory features and structural features from the preprocessed image; carrying out cross-modal fusion on the extracted multi-scale features; calculating an activity quantitative score based on the fused comprehensive features; and outputting a visual report of the quantitative score and the evaluation conclusion. The system comprises an image acquisition module, an image preprocessing module, a feature extraction module, a feature fusion module, an activity evaluation module and a result output module. According to the technical scheme, objective quantitative evaluation of thyroid eye disease activity can be achieved, the consistency and repeatability of evaluation results are improved, remote image acquisition and automatic analysis are supported, and the space-time limitation of traditional clinical evaluation is broken through.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method and system for assessing the activity of thyroid eye disease based on eye images. Background Technology

[0002] Thyroid ophthalmopathy is a chronic inflammatory orbital disease closely related to autoimmune thyroid diseases. Its clinical progression is typically divided into active and inactive phases, with fundamental differences in pathological characteristics, treatment response, and intervention strategies between the two phases. This disease has a high incidence rate globally, especially among middle-aged women. Failure to promptly identify and effectively intervene in the active phase can easily lead to visual impairment, proptosis, and even permanent ocular dysfunction. Therefore, accurate and dynamic assessment of disease activity is crucial for developing individualized treatment plans, optimizing clinical decision-making pathways, and improving patient prognosis.

[0003] Among these, disease activity assessment based on ocular characteristics is a core component of current clinical practice. The mainstream approach relies on clinical activity scoring systems, which make a binary judgment based on observations of several clinical signs such as spontaneous retrobulbar pain, pain during eye movement, eyelid redness and swelling, and conjunctival hyperemia. However, such methods face multiple challenges in practical application: First, the scoring results are discrete integer values, making it difficult to reflect the continuous evolution of disease activity; second, the assessment process is highly dependent on the physician's subjective experience and the patient's immediate condition, easily introducing inter- and intra-observer variability, affecting the consistency and reproducibility of the results; third, existing scoring standards are mostly derived from Western population cohorts, and their applicability in the Chinese population has not been fully validated, potentially leading to biases due to racial differences; furthermore, because thyroid ophthalmopathy has a prolonged course and long follow-up period, and often requires comprehensive assessment in conjunction with thyroid function status, traditional scoring methods are limited by spatial and temporal conditions, making it difficult to achieve high-frequency, remote dynamic monitoring.

[0004] Current technology lacks a method for accurately quantifying the activity of thyroid eye disease based solely on objective ocular image data, without relying on manual clinical scoring. This deficiency hinders real-time, non-invasive, and standardized tracking of disease activity in clinical practice, restricting the in-depth application of precision medicine in this field. Therefore, there is an urgent need to develop an automated and intelligent assessment method and system based on ocular images. By deeply mining the inflammatory and structural features contained in the images, this method can achieve highly sensitive and specific discrimination of disease activity, thereby providing clinicians with a quantifiable, reproducible, and remotely deployable decision support tool. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to provide a method and system for assessing the activity of thyroid eye disease based on ocular images, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In one aspect, a thyroid eye disease activity assessment system based on ocular images, comprising the following components: an image acquisition module for acquiring raw image data containing the ocular region; an image preprocessing module connected to the image acquisition module for standardizing the raw image data to eliminate illumination differences and geometric distortions; a feature extraction module connected to the image preprocessing module for extracting multi-scale inflammatory features and structural features from the preprocessed image; a feature fusion module connected to the feature extraction module for cross-modal fusion of multi-scale features to generate a comprehensive feature representation; an activity assessment module connected to the feature fusion module for calculating a quantitative score of disease activity based on the comprehensive feature representation; and a result output module connected to the activity assessment module for presenting the quantitative score and assessment conclusion in a visual form.

[0007] Preferably, the image acquisition module supports eye image data acquired by various imaging devices, including but not limited to digital SLR cameras, smartphone cameras, and dedicated ophthalmic imaging equipment; the module is equipped with an autofocus mechanism and a ring LED fill light system to ensure that images with a resolution of not less than 1920×1080 pixels are obtained within a shooting distance of 300mm to 500mm; the module also integrates an image quality evaluation unit, which provides real-time feedback on the acquisition quality based on quantitative indicators of image sharpness, contrast, and noise level.

[0008] Furthermore, the image preprocessing module performs the following operations: color space conversion, converting the original RGB image to Lab color space to separate brightness and color information; illumination correction, using an image enhancement algorithm based on Retinex theory to eliminate the influence of non-uniform illumination; image registration, aligning images acquired at different time points to a unified coordinate system through affine transformation; and region of interest extraction, using a cascaded convolutional neural network to locate and crop a standardized region containing the eyelids, conjunctiva, and cornea.

[0009] In addition, the feature extraction module includes two sub-networks that work in parallel: the inflammatory feature extraction sub-network adopts a deep residual network architecture, whose input is a preprocessed eye image, and extracts visual features from low-level edge texture to high-level semantics layer by layer through 5 residual blocks; the structural feature extraction sub-network adopts a U-Net architecture to achieve pixel-level segmentation and accurately quantifies morphological parameters such as eyelid retraction distance, eyeball protrusion and conjunctival vascular density.

[0010] Preferably, the feature fusion module uses an attention mechanism to achieve cross-modal feature interaction, specifically including: constructing feature mapping matrices for inflammatory features and structural features respectively; determining the correlation strength of the two features in the spatial dimension by calculating the cross-attention weight matrix; and performing weighted fusion of features based on attention weights to generate a comprehensive feature vector with a dimension of 512.

[0011] Furthermore, the activity assessment module quantifies disease activity based on a comprehensive feature vector, with its core being a pre-trained multilayer perceptron model. This model contains three fully connected layers, uses the ReLU activation function, and the output layer uses the Sigmoid function to map the calculation results to continuous values ​​between 0 and 1, which directly represent the degree of activity of thyroid eye disease.

[0012] In addition, the results output module generates an assessment report containing the following: a quantitative score of disease activity and its position markings on a color scale bar; a clinical staging suggestion corresponding to the score; a visualization of key features, including a heat map of inflammatory areas and a structural parameter comparison chart; and a historical score trend curve for tracking the disease evolution process.

[0013] To achieve the above objectives, the present invention provides the following technical solution: On the other hand, a method for assessing the activity of thyroid eye disease based on ocular images, the method comprising the following steps: Step S110, acquiring raw image data containing the ocular region; Step S120, performing standardized preprocessing on the raw image data; Step S130, extracting multi-scale inflammatory features and structural features from the preprocessed image; Step S140, performing cross-modal fusion of the extracted multi-scale features; Step S150, calculating a quantitative score of disease activity based on the fused comprehensive features; Step S160, outputting a visual report containing the quantitative score and assessment conclusion.

[0014] Preferably, the standardized preprocessing in step S120 includes four sub-steps: color space conversion, illumination correction, image registration, and region of interest extraction. Color space conversion uses the CIE Lab standard, where the L component represents luminance, the a component represents red-green hue, and the b component represents yellow-blue hue. Illumination correction is achieved by solving an optimization problem of $$L(x,y)=R(x,y)×I(x,y)$$, where R(x,y) is the reflection component and I(x,y) is the illumination component. Image registration uses a SIFT-based feature point matching algorithm, with the registration error controlled within 2 pixels. Region of interest extraction is achieved through a trained target detection network, outputting a standard region with a fixed size of 512×512 pixels.

[0015] Furthermore, the feature extraction process in step S130 handles inflammatory features and structural features respectively. Inflammatory feature extraction uses a ResNet50 network pre-trained on ImageNet, removing its fully connected layers and retaining the convolutional and pooling layers, and generating a 2048-dimensional feature vector through global average pooling. Structural feature extraction first locates the eyelid edge, corneal contour and conjunctival region through a semantic segmentation network, and then calculates the eyelid retraction distance (measures the vertical distance between the upper eyelid edge and the upper edge of the cornea), eyeball protrusion (calculates the distance between the corneal apex and the orbital rim based on three-dimensional reconstruction technology), and conjunctival vascular density (statistically calculates the percentage of vascular pixels per unit area through a vascular segmentation algorithm).

[0016] Furthermore, the feature fusion in step S140 employs a bilinear attention pooling mechanism, specifically implemented as follows: Let the inflammatory feature matrix be $F_i∈R^{H×W×C_i}$, and the structural feature matrix be $F_s∈R^{H×W×C_s}$, and the bilinear features are obtained through outer product calculation. The dimensionality is then reduced to a fixed dimension through global average pooling; this fusion mechanism preserves the second-order statistical information between the two features, enhancing the model's ability to perceive subtle pathological changes.

[0017] Preferably, the activity score calculation in step S150 is based on a deep neural network model, which is trained with a large amount of labeled data. The training data includes 3,000 cases of thyroid eye disease confirmed by clinical experts. Each case includes eye images at multiple time points and corresponding clinical activity scores. The model training uses the mean squared error loss function, the optimizer is Adam, the initial learning rate is set to 0.001, and the training cycle is 100 rounds. The trained model achieves a correlation coefficient of 0.92 with the clinical expert evaluation results on the independent test set.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. It has achieved an objective quantitative assessment of the activity of thyroid eye disease, transforming traditional discrete clinical scores into continuous values ​​that can accurately reflect subtle changes in the degree of disease activity;

[0020] 2. By replacing subjective human judgment with automated image analysis, the variation between and within observers is eliminated, ensuring the consistency and repeatability of the evaluation results;

[0021] 3. Training and validation based on ocular image data of the Chinese population avoids assessment bias caused by racial differences and improves adaptability in clinical applications in China;

[0022] 4. It supports remote image acquisition and automated analysis, breaking through the time and space limitations of traditional clinical assessment and realizing high-frequency dynamic monitoring of disease activity;

[0023] 5. It integrates multi-dimensional information from inflammatory and structural features, providing a more comprehensive and accurate basis for judging disease activity than a single feature. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical architecture of the method and system for assessing the activity of thyroid eye disease based on eye images proposed in this invention.

[0025] Figure 2 This is a schematic diagram of the core principle framework for the extraction and fusion of multi-scale inflammatory features and structural features in this invention.

[0026] Figure 3 This is a flowchart illustrating the logical flow of the image preprocessing and feature extraction stages in this invention.

[0027] Figure 4 This is a logical flowchart of the feature fusion and activity evaluation stages in this invention.

[0028] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of image acquisition, processing and result output in this invention. Detailed Implementation

[0029] 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.

[0030] Example 1

[0031] In the thyroid eye disease clinic of a tertiary-level Class A hospital's ophthalmology department, a thyroid eye disease activity assessment system based on ocular images is comprehensively conducting data collection and processing. (See also...) Figure 1 This system includes an image acquisition module, an image preprocessing module, a feature extraction module, a feature fusion module, an activity assessment module, and a result output module. The image acquisition module uses dedicated ophthalmic imaging equipment equipped with an autofocus mechanism and a ring LED illumination system. It acquires raw image data of the eye region with a resolution of 1920 x 1080 pixels within a shooting distance of 300 mm to 500 mm. The integrated image quality assessment unit monitors image sharpness, contrast, and noise levels in real time. When sharpness falls below 0.8, contrast below 0.6, or signal-to-noise ratio falls below 30 dB, it automatically prompts for re-acquisition to ensure the reliability of the input data.

[0032] The image preprocessing module standardizes the acquired raw image data. (See also...) Figure 3This module first performs a color space conversion, transforming the original RGB image to the Lab color space, where the L component represents luminance information, the a component represents red-green hue information, and the b component represents yellow-blue hue information. Next, illumination correction is performed using an image enhancement algorithm based on Retinex theory, eliminating the influence of non-uniform illumination by solving an optimization problem between the reflection and illumination components. The image registration submodule aligns images acquired at different time points to a unified coordinate system through affine transformation, employing a SIFT feature point matching algorithm to strictly control the registration error within 2 pixels. Finally, a cascaded convolutional neural network is used to locate and crop a standardized region containing the eyelids, conjunctiva, and cornea, outputting a region of interest with a fixed size of 512 x 512 pixels.

[0033] The feature extraction module extracts multi-scale inflammatory and structural features from the preprocessed image. See also Figure 2 This module comprises two sub-networks operating in parallel. The inflammatory feature extraction sub-network employs a deep residual network architecture, taking a preprocessed eye image as input and extracting visual features layer by layer through five residual blocks. The first residual block extracts low-level edge texture features, including the orientation patterns of conjunctival vessels and the contour information of the eyelid edges. The second residual block extracts mid-level semantic features, identifying the color distribution and degree of edema in the conjunctival hyperemia area. The subsequent three residual blocks extract high-level semantic features, capturing the stereoscopic information of eyelid swelling and the degree of restriction in eye movement. This sub-network uses a ResNet50 network pre-trained on ImageNet, removing its fully connected layers while retaining convolutional and pooling layers, and generating a 2048-dimensional feature vector through global average pooling.

[0034] The structural feature extraction subnetwork employs a U-Net architecture to achieve pixel-level segmentation and accurately quantify morphological parameters. This network first extracts multi-scale features through encoder path downsampling, and then restores spatial resolution through decoder path upsampling. The network output includes three segmentation masks: an eyelid edge mask for calculating eyelid retraction distance, i.e., measuring the vertical distance between the upper eyelid edge and the upper edge of the cornea, with an accuracy of 0.1 mm; a corneal contour mask for calculating eyeball protrusion, calculating the distance between the corneal apex and the orbital rim based on 3D reconstruction technology, with a measurement error of less than 0.5 mm; and a conjunctival region mask for calculating conjunctival vessel density, statistically analyzing the percentage of vessel pixels per unit area using a vessel segmentation algorithm, achieving a resolution of 100 sampling points per square millimeter.

[0035] The feature fusion module fuses multi-scale features across modalities to generate a comprehensive feature representation. See also Figure 4This module employs an attention mechanism to achieve cross-modal feature interaction. First, feature mapping matrices are constructed for inflammatory and structural features respectively. The inflammatory feature matrix has 2048 dimensions, while the structural feature matrix includes parameters for eyelid retraction distance, eyeball protrusion, and conjunctival vascular density. The correlation strength between the two features in the spatial dimension is determined by calculating the cross-attention weight matrix. Specifically, a bilinear attention pooling mechanism is used. Let the inflammatory feature matrix be F_i and the structural feature matrix be F_s. The bilinear feature F_b is obtained through outer product calculation, and then reduced to a fixed 512-dimensional dimension through global average pooling. This fusion mechanism preserves the second-order statistical information between the two features, enhancing the model's ability to perceive subtle pathological changes.

[0036] The disease activity assessment module calculates a quantitative score for disease activity based on a comprehensive feature representation. The core of this module is a pre-trained multilayer perceptron model containing three fully connected layers, using the ReLU activation function. The output layer employs a sigmoid function to map the calculation results to continuous values ​​between 0 and 1. This model was trained on 3000 clinically confirmed cases of thyroid ophthalmopathy, each case including ocular images at multiple time points and corresponding clinical activity scores. The model training used a mean squared error loss function, with Adam as the optimizer, an initial learning rate of 0.001, and 100 training iterations. The trained model achieved a correlation coefficient of 0.92 with the clinical expert assessment results on an independent test set. The score output is a continuous numerical value: 0 to 0.3 represents inactive periods, 0.3 to 0.7 represents mild to moderate active periods, and 0.7 to 1.0 represents severe active periods.

[0037] The results output module presents the quantitative scores and evaluation conclusions in a visual format. See also Figure 5 This module generates an assessment report containing the following: a quantitative score of disease activity and its position on a color-coded scale bar, with the color gradient from green to red corresponding to scores from 0 to 1; clinical staging recommendations corresponding to the score, including observation for the inactive phase, drug treatment recommendations for the mildly active phase, and surgical intervention recommendations for the severely active phase; visualization of key features, including a heatmap of the inflammation area and a structural parameter comparison chart. The heatmap displays the spatial distribution of inflammation severity in pseudo-color, and the structural parameter comparison chart shows the difference between the current measurement value and the normal reference value in bar chart form; and a historical score trend curve to track the disease evolution process, supporting backtracking of up to 12 months of historical data.

[0038] In specific clinical applications, after a patient enters the examination room, the physician operates the image acquisition module to obtain standard frontal and lateral views of both eyes. The system completes image quality assessment within 10 seconds, and if the quality meets the standards, it automatically triggers the preprocessing process. The preprocessing stage takes approximately 15 seconds, including color space conversion, illumination correction, image registration, and region of interest extraction. The feature extraction stage processes inflammatory and structural features in parallel, taking approximately 20 seconds. The feature fusion and activity assessment stage takes approximately 5 seconds, ultimately generating a complete assessment report within 1 minute. This system achieves an objective quantitative assessment of thyroid eye disease activity, transforming traditional discrete clinical scores into continuous values ​​that accurately reflect subtle changes in disease activity. By replacing subjective human judgment with automated image analysis, it eliminates inter- and intra-observer variability, ensuring the consistency and reproducibility of assessment results. Training and validation based on ocular image data from the Chinese population avoids assessment bias caused by racial differences, improving its adaptability in clinical applications in China.

[0039] Example 2

[0040] In the telemedicine platform of a regional medical center, a thyroid eye disease activity assessment system based on ocular images supports disease screening and follow-up management in primary healthcare institutions. The image acquisition module is equipped with a smartphone camera adapter interface, guiding primary care physicians or patients to acquire ocular images themselves through a dedicated image acquisition application. The application has a built-in shooting guidance function, ensuring images meet standard requirements through real-time face detection and eye positioning. When the shooting distance is detected to be beyond 300mm to 500mm or the image resolution is lower than 1920 x 1080 pixels, the application automatically prompts for adjustment of the shooting position.

[0041] The image preprocessing module is optimized for the characteristics of images captured by mobile devices. Color space conversion uses a mobile-optimized Lab conversion algorithm, reducing computational complexity by 40% while maintaining color accuracy. The illumination correction algorithm is enhanced for the mixed lighting environments common on mobile devices, effectively eliminating color casts caused by a mixture of fluorescent and natural light. Image registration employs a lightweight feature point detection algorithm, reducing computational resource usage by 60% while maintaining registration accuracy. Region of interest extraction uses a compressed convolutional neural network model, reducing the model size from 50 megabytes to 15 megabytes, making it more suitable for deployment and operation on mobile devices.

[0042] The feature extraction module runs on a cloud server, receiving preprocessed images via encrypted data transmission. The inflammatory feature extraction sub-network employs a residual network compressed using knowledge distillation technology. The teacher network is a full ResNet50 model, while the student network is a simplified ResNet18 model, improving inference speed by 2.5 times while maintaining 95% feature extraction accuracy. The structural feature extraction sub-network uses an improved U-Net architecture, introducing depthwise separable convolutions to reduce the number of parameters, reducing model computation by 45% while maintaining segmentation accuracy. The feature fusion module employs a hierarchical attention mechanism, performing coarse-grained fusion first at low-level feature layers and then fine-grained fusion at high-level feature layers, ensuring both fusion effectiveness and controlled computational complexity.

[0043] The activity assessment module is deployed on a local server within the healthcare institution, receiving fused feature vectors from the cloud. This module employs a lightweight multilayer perceptron model with only two fully connected layers, reducing inference time from 50 milliseconds to 20 milliseconds while maintaining assessment accuracy. The model is specifically optimized for telemedicine scenarios, enhancing robustness to image quality fluctuations and maintaining stable assessment performance even with slightly blurred or unevenly lit input images. In addition to numerical results, the scoring output includes a confidence level indicator; if the confidence level falls below 0.85, it is recommended to re-acquire images for assessment.

[0044] The results output module generates a simplified assessment report suitable for telemedicine scenarios. The report focuses on core assessment results, including current activity scores, trends compared to the previous assessment, and score-based management recommendations. The report is pushed to physicians and patients through the healthcare institution's patient portal or mobile application, and can be viewed and downloaded online. The system also provides batch processing capabilities, allowing primary healthcare institutions to upload image data from multiple patients at once. The system processes all data and generates a summary report within two hours, assisting healthcare institutions in conducting population screening and epidemiological analysis.

[0045] This system supports long-term follow-up management of thyroid eye diseases. Each assessment result is automatically stored in the patient's electronic health record, forming a complete record of disease progression. Physicians can view the comparison of assessment results at any point in time through the online platform. The system automatically generates disease progression trend charts, visually displaying treatment effectiveness or the natural progression of the disease. When the system detects that a patient's activity score increases by more than 0.2 twice consecutively or reaches 0.7 or higher in a single score, it automatically triggers a high-risk warning, notifying the responsible physician via SMS and system messages for timely intervention. This assessment mode, based on remote image acquisition and automated analysis, breaks through the spatial and temporal limitations of traditional clinical assessments, enabling high-frequency dynamic monitoring of disease activity, and is particularly suitable for the management of chronic disease patients requiring long-term follow-up.

[0046] 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 system for assessing thyroid ophthalmopathy activity based on ocular images, characterized in that, The system includes the following components: The image acquisition module is used to acquire raw image data containing the eye region. This module supports eye image data acquired by various imaging devices and is equipped with an autofocus mechanism and a ring LED fill light system. It can obtain images with a resolution of no less than 1920 x 1080 pixels within a shooting distance of 300 mm to 500 mm. This module also integrates an image quality evaluation unit, which provides real-time feedback on the acquisition quality based on quantitative indicators of image sharpness, contrast and noise level. The image preprocessing module, connected to the image acquisition module, is used to standardize the raw image data to eliminate illumination differences and geometric distortions. This module performs color space conversion operations, converting the original RGB image to the Lab color space to separate brightness and color information. It also performs illumination correction operations, using an image enhancement algorithm based on Retinex theory to eliminate the effects of non-uniform illumination. Furthermore, it performs image registration operations, aligning images acquired at different time points to a unified coordinate system through affine transformations. Finally, it performs region of interest extraction operations, using a cascaded convolutional neural network to locate and crop the standardized region containing the eyelids, conjunctiva, and cornea. The feature extraction module, connected to the image preprocessing module, is used to extract multi-scale inflammatory and structural features from the preprocessed image. This module contains two sub-networks that work in parallel. The inflammatory feature extraction sub-network adopts a deep residual network architecture. Its input is the preprocessed eye image. It extracts visual features from low-level edge texture to high-level semantics layer by layer through 5 residual blocks. The structural feature extraction sub-network adopts a U-Net architecture to achieve pixel-level segmentation and accurately quantifies morphological parameters such as eyelid retraction distance, eyeball protrusion, and conjunctival vascular density. The feature fusion module, connected to the feature extraction module, is used to perform cross-modal fusion of multi-scale features to generate a comprehensive feature representation. This module uses an attention mechanism to realize cross-modal feature interaction, constructs feature mapping matrices for inflammatory features and structural features respectively, determines the correlation strength of the two features in the spatial dimension by calculating the cross-attention weight matrix, and performs weighted fusion of features based on attention weights to generate a comprehensive feature vector with a dimension of 512. The activity assessment module, connected to the feature fusion module, is used to calculate a quantitative score of disease activity based on the comprehensive feature representation. The core of this module is a pre-trained multilayer perceptron model, which contains three fully connected layers and uses the ReLU activation function. The output layer uses the Sigmoid function to map the calculation results to a continuous value between 0 and 1, which directly represents the degree of activity of thyroid eye disease. The results output module, connected to the activity assessment module, is used to present the quantitative scores and assessment conclusions in a visual form. This module generates an assessment report that includes the quantitative score of disease activity and its position markings on the color scale bar, clinical staging suggestions corresponding to the score, visualization of key features, and historical score trend curves.

2. The thyroid ophthalmopathy activity assessment system based on ocular images according to claim 1, characterized in that, In the image preprocessing module, the image registration operation adopts a SIFT feature point-based matching algorithm, and the registration error is controlled within 2 pixels; the region of interest extraction operation outputs a standard region with a fixed size of 512 by 512 pixels.

3. The thyroid ophthalmopathy activity assessment system based on ocular images according to claim 1, characterized in that, In the feature extraction module, the inflammation feature extraction sub-network uses a ResNet50 network pre-trained on ImageNet. After removing its fully connected layers, the convolutional and pooling layers are retained, and a 2048-dimensional feature vector is generated through global average pooling. The structural feature extraction subnetwork locates the eyelid edge, corneal contour, and conjunctival region through a semantic segmentation network, and then calculates the eyelid retraction distance, eyeball protrusion, and conjunctival vascular density.

4. The thyroid ophthalmopathy activity assessment system based on ocular images according to claim 1, characterized in that, The feature fusion module employs a bilinear attention pooling mechanism. Let the inflammatory feature matrix be $F_i∈R^{H×W×C_i}$, and the structural feature matrix be $F_s∈R^{H×W×C_s}$. The bilinear features are obtained through outer product calculation. Then, global average pooling is used to reduce the dimensionality to a fixed dimension.

5. The thyroid ophthalmopathy activity assessment system based on ocular images according to claim 1, characterized in that, The model for the activity assessment module was trained using 3,000 clinically confirmed cases of thyroid eye disease. Each case included ocular images at multiple time points and corresponding clinical activity scores. The model was trained using the mean squared error loss function, with Adam selected as the optimizer, an initial learning rate of 0.001, and a training cycle of 100 rounds.

6. The thyroid ophthalmopathy activity assessment system based on ocular images according to claim 1, characterized in that, In the result output module, the key feature visualization includes a heat map of the inflammatory area and a structural parameter comparison chart. The heat map displays the spatial distribution of the degree of inflammation in pseudo-color, and the structural parameter comparison chart displays the difference between the current measurement value and the normal reference value in the form of a bar chart.

7. A method for assessing thyroid ophthalmopathy activity based on ocular images, characterized in that, The method includes the following steps: S110 acquires raw image data including the eye area. Through an imaging device equipped with an autofocus mechanism and a ring LED fill light system, it obtains images with a resolution of no less than 1920 x 1080 pixels within a shooting distance of 300 mm to 500 mm, and evaluates the acquisition quality in real time based on quantitative indicators of image sharpness, contrast and noise level. S120 performs standardized preprocessing on the original image data, including color space conversion, illumination correction, image registration, and region of interest extraction. The color space conversion converts the original RGB image to the Lab color space. Illumination correction uses an image enhancement algorithm based on Retinex theory. Image registration uses affine transformation to align images acquired at different time points to a unified coordinate system. Region of interest extraction uses a cascaded convolutional neural network to locate and crop standardized regions containing eyelids, conjunctiva, and cornea. S130 extracts multi-scale inflammatory and structural features from the preprocessed image. The inflammatory feature extraction uses a deep residual network architecture, which extracts visual features from low-level edge texture to high-level semantics layer by layer through 5 residual blocks. The structural feature extraction uses a U-Net architecture to achieve pixel-level segmentation and accurately quantifies morphological parameters such as eyelid retraction distance, eyeball protrusion and conjunctival vascular density. S140: The extracted multi-scale features are fused across modalities. An attention mechanism is used to achieve cross-modal feature interaction. Feature mapping matrices are constructed for inflammatory features and structural features respectively. By calculating the cross-attention weight matrix, the correlation strength of the two features in the spatial dimension is determined. Based on the attention weight, the features are weighted and fused to generate a comprehensive feature vector with a dimension of 512. S150 calculates a quantitative score of disease activity based on the fused comprehensive features. It uses a pre-trained multilayer perceptron model containing three fully connected layers, uses the ReLU activation function, and the output layer uses the Sigmoid function to map the calculation results to continuous values ​​between 0 and 1. S160 outputs a visual report containing quantitative scores and assessment conclusions, generating quantitative scores for disease activity and their position markers on color scale bars, clinical staging suggestions corresponding to the scores, visualization of key features, and historical score trend curves.

8. The method for assessing thyroid ophthalmopathy activity based on ocular images according to claim 7, characterized in that, The standardization preprocessing in step S120 includes color space conversion using the CIE Lab standard, where the L component represents luminance, the a component represents red-green hue, and the b component represents yellow-blue hue; illumination correction is achieved by solving an optimization problem of $L(x,y)=R(x,y)×I(x,y)$, where R(x,y) is the reflection component and I(x,y) is the illumination component.

9. The method for assessing thyroid ophthalmopathy activity based on ocular images according to claim 7, characterized in that, The feature extraction process in step S130 handles inflammatory features and structural features respectively; Inflammatory feature extraction uses a ResNet50 network pre-trained on ImageNet. After removing its fully connected layers, convolutional and pooling layers are retained, and a 2048-dimensional feature vector is generated by global average pooling. Structural feature extraction first locates the eyelid edge, corneal contour and conjunctival region through a semantic segmentation network, and then calculates the eyelid retraction distance, eyeball protrusion and conjunctival vascular density.

10. The method for assessing thyroid ophthalmopathy activity based on ocular images according to claim 7, characterized in that, The feature fusion in step S140 employs a bilinear attention pooling mechanism. Let the inflammatory feature matrix be $F_i∈R^{H×W×C_i}$ and the structural feature matrix be $F_s∈R^{H×W×C_s}$. The bilinear features are obtained through outer product calculation. Then, global average pooling is used to reduce the dimensionality to a fixed dimension.

Images