A lightweight assisted diagnostic and treatment system and method for thyroid-related ophthalmopathy based on ultrasound imaging.
By using multi-structure quantization and normalization processing based on ultrasound images, the problem of insufficient diagnosis in the image-assisted diagnosis and treatment of thyroid-related ophthalmopathy is solved. It achieves efficient and low-cost multi-structure quantization and diagnosis, which is convenient for use in primary healthcare institutions and improves the accuracy and comparability of diagnosis and treatment.
Patent Information
- Application Number
- CN202611433105.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-16
AI Technical Summary
Existing imaging-assisted diagnosis and treatment technologies for thyroid-related ophthalmopathy mainly suffer from problems such as low detection rate of atypical subtypes, lack of comparability and reproducibility of results due to equipment and individual differences, insufficient awareness in primary healthcare institutions, and high examination costs, resulting in low early diagnosis rates and missed opportunities for optimal treatment.
Based on ultrasound images, by simultaneously quantifying the oculomotor muscle, orbital fat, and optic nerve, and using the vitreous features within the eyeball as an internal reference benchmark, normalization processing is performed to construct a joint feature vector, thereby achieving multi-structure quantification and assisted diagnosis and treatment, reducing costs, and facilitating follow-up.
It improves the comprehensiveness and accuracy of diagnosis and treatment of thyroid-related ophthalmopathy, reduces the rate of missed diagnosis, enhances the comparability and reproducibility of results, is applicable to outpatient clinics and primary care institutions, is convenient and low-cost, and meets clinical needs.
Smart Images

Figure CN122966645B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing and computer-aided diagnosis and treatment technology, specifically to a lightweight auxiliary diagnosis and treatment system and method for thyroid-related ophthalmopathy based on ultrasound images. Background Technology
[0002] Thyroid-associated ophthalmopathy (TAO, also known as Graves' ophthalmopathy) is an autoimmune disease affecting the extraocular muscles, orbital fat, optic nerve, and microcirculation within the orbit. Accurate imaging assessment is crucial for diagnosis, classification, and follow-up. Current image-assisted diagnosis and treatment have evolved from single-structure, single-modality approaches to multi-structure, multi-modality approaches.
[0003] Thyroid-associated ophthalmopathy is an autoimmune orbital disease whose pathological changes affect the orbital structures behind the eyeball, namely hypertrophy of the rectus muscle, orbital fat hyperplasia, and optic nerve involvement. It is an orbital-level disease rather than an ocular-level disease.
[0004] In terms of pathological features, thyroid-associated ophthalmopathy can affect multiple orbital structures simultaneously or separately, and its dynamic changes with the course of the disease and treatment require simultaneous quantification of multiple structures and longitudinal follow-up comparison. In contrast, the identification of ordinary intraocular diseases is mainly based on lesion localization and type judgment, without the need for cross-device and cross-time comparable quantification of multiple orbital structures.
[0005] Current image-assisted diagnosis and treatment techniques for thyroid-associated ophthalmopathy mainly include methods based on computed tomography (CT), magnetic resonance imaging (MRI), and clinical activity score (CAS) prediction based on facial photographs. However, existing image-assisted diagnosis and treatment techniques for thyroid-associated ophthalmopathy have the following main shortcomings: First, most protocols focus only on a single anatomical structure (such as measuring only the volume of extraocular muscles), while the pathological changes in thyroid-associated ophthalmopathy can occur simultaneously or separately in the rectus abdominis muscle, orbital fat, and optic nerve, resulting in a low detection rate for atypical subtypes and a risk of missed diagnosis.
[0006] Secondly, the image grayscale values based on computed tomography (CT) or magnetic resonance imaging (MRI) are significantly affected by the equipment model and scanning parameters, and often use indicators with large individual differences, such as orbital cavity volume, for normalization, resulting in a lack of comparability and reproducibility of results between different medical institutions.
[0007] Furthermore, most patients and even primary healthcare institutions lack sufficient awareness of thyroid-associated ophthalmopathy. In addition, the gold standard examinations such as MRI have low availability and high cost in primary healthcare institutions, which prevents some patients from seeking medical attention in the early stages. They are only identified when symptoms become obvious, at which point they may have already experienced severe symptoms such as vision loss or even blindness, missing the best time for intervention and treatment, and ultimately causing irreversible visual impairment.
[0008] Therefore, there is an urgent need for a new, low-cost, and convenient auxiliary diagnosis and treatment solution for thyroid-related ophthalmopathy that can simultaneously quantify multiple key orbital structures, eliminate device and individual differences, and provide convenient and efficient treatment. Summary of the Invention
[0009] The purpose of this invention is to provide a lightweight auxiliary diagnostic and treatment system and method for thyroid-related ophthalmopathy based on ultrasound images, so as to solve the problems mentioned in the background art.
[0010] The main design concept of this invention is as follows: To address the aforementioned technical challenges, this invention aims to provide a lightweight auxiliary diagnostic and treatment system and method for thyroid-associated ophthalmopathy based on ultrasound images. This method uses ocular ultrasound images as a foundation to simultaneously and quantitatively analyze multiple key orbital anatomical structures (rectus muscle, orbital fat, and optic nerve). It utilizes stable vitreous features within the eyeball as an internal reference benchmark, performing ratio normalization on size features, mean grayscale features, and grayscale standard deviation features based on vitreous features. Dimensionless morphological features and texture statistics are directly retained, ultimately forming an interpretable joint feature vector that is input into a classifier model. This enables a more comprehensive, robust, and interpretable auxiliary diagnosis and treatment of thyroid-associated ophthalmopathy, while also offering advantages such as being radiation-free, low-cost, convenient, and suitable for follow-up.
[0011] For thyroid-associated ophthalmopathy, the rectus muscle, orbital fat, and optic nerve are the pathologically relevant structures that this invention focuses on quantifying; the vitreous body is used to establish the patient's own internal benchmark because it is not one of the main target structures of the disease mentioned above.
[0012] Therefore, this invention is not simply "segmenting an organ," but rather transforming an intraocular structure that is not typically used as a criterion for structural abnormalities in thyroid-associated ophthalmopathy into a reference system for patient-standardized analysis of other structures.
[0013] To achieve the above objectives, the present invention provides the following technical solution: On one hand, the present invention provides a lightweight auxiliary diagnosis and treatment system for thyroid-related ophthalmopathy based on ultrasound images, comprising the following modules: The input module is used to acquire and input the patient's ocular ultrasound images; The region segmentation module is used to segment the rectus abdominis region, orbital fat region, optic nerve region, and vitreous region based on ocular ultrasound images; The feature extraction module is used to extract features of the rectus abdominis muscle in the rectus abdominis muscle region, features of the orbital fat in the orbital fat region, features of the optic nerve in the optic nerve region, and features of the vitreous body in the vitreous body region, respectively. The normalization module is used to normalize the features of the oculomotor rectus muscle, orbital fat and optic nerve respectively, using the vitreous features as the internal reference benchmark, to obtain normalized features of multiple orbital structures. The feature vector module is used to construct joint feature vectors based on normalized features of multiple orbital structures; The classification and diagnosis module, based on joint feature vectors, obtains and outputs auxiliary diagnosis and treatment results for thyroid-related eye diseases through a classifier model. The normalization module includes: The rectus abdominis feature quantization unit is used to obtain the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis region, and relative gray level standard deviation of the rectus abdominis region, using the vitreous features as an internal reference benchmark. The orbital fat feature quantification unit is used to obtain the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region with vitreous features as an internal reference. The optic nerve feature quantification unit is used to obtain the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of optic nerve features using vitreous features as an internal reference.
[0014] More preferably, the rectus abdominis feature quantification unit is used to obtain the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis region, and relative gray level standard deviation of the rectus abdominis region, using vitreous features as an internal reference benchmark. Specifically, this is achieved through the following methods: Relative maximum thickness ;in, For maximum thickness, The anteroposterior diameter is a characteristic feature of the vitreous body; relative average thickness ;in, Average thickness; relative cross-sectional area ;in, For cross-sectional area, The area of the ellipse characteristic of the vitreous body; Relative major axis length ;in, The length of the major axis. The transverse diameter is a characteristic feature of the vitreous body; Relative minor axis length ;in, The length of the minor axis; The relative average gray level of the rectus abdominis region ;in, The average gray level of the rectus abdominis muscle region. The average gray level of the vitreous region; Relative gray standard deviation of the rectus abdominis muscle region ;in, The standard deviation of grayscale values in the rectus abdominis muscle region. denoted as the standard deviation of grayscale in the vitreous region.
[0015] More preferably, the orbital fat feature quantification unit is used to obtain the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region using vitreous features as an internal reference benchmark, specifically obtained through the following methods: relative area ;in, For area; relative volume ;in, For volume; Relative average gray level of orbital fat region ;in, The average gray level of the orbital fat region; Relative gray standard deviation of orbital fat region ;in, The gray standard deviation represents the orbital fat region.
[0016] More preferably, the optic nerve feature quantification unit is used to obtain the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features using vitreous features as an internal reference benchmark, specifically through the following methods: relative average diameter ;in, The average diameter; Relative maximum diameter ;in, The maximum diameter; Relative minimum diameter ;in, Minimum diameter; Relative optic nerve sheath thickness ;in, The thickness of the nerve sheath; Relative average gray level ;in, This represents the average gray level of the optic nerve region.
[0017] On the other hand, the present invention provides a lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images, comprising the following steps: Acquire ultrasound images of the patient's eyes; Based on ocular ultrasound images, the rectus abdominis region, orbital fat region, optic nerve region, and vitreous region were segmented. Features of the rectus abdominis muscle in the rectus abdominis muscle region, features of the orbital fat in the orbital fat region, features of the optic nerve in the optic nerve region, and features of the vitreous body in the vitreous body region were extracted respectively. Using vitreous features as an internal reference, the features of the rectus abdominis muscle, orbital fat, and optic nerve were normalized to obtain normalized features of multiple orbital structures, including: Using vitreous features as an internal reference, the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus muscle region, and relative gray level standard deviation of the rectus muscle region are obtained. Using vitreous features as an internal reference, the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region were obtained. Using vitreous features as an internal reference, the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features were obtained. Construct a joint feature vector based on the normalized features of multiple orbital structures; Based on joint feature vectors, the auxiliary diagnosis and treatment results of thyroid-related ophthalmopathy are obtained through a classifier model.
[0018] More preferably, the features of the rectus abdominis muscle include: maximum thickness, average thickness, cross-sectional area, long axis length, short axis length, average gray level of the rectus abdominis muscle region, standard deviation of gray level of the rectus abdominis muscle region, roundness, ratio of long axis to short axis, and ratio of maximum thickness to average thickness. The orbital fat features include: area, volume, average gray level of the orbital fat region, standard deviation of gray level of the orbital fat region, contrast, energy, entropy, and homogeneity. The optic nerve features include: average diameter, maximum diameter, minimum diameter, optic nerve sheath thickness, and average gray level of the optic nerve region; The characteristics of the vitreous body include: anteroposterior diameter, transverse diameter, elliptical area, average gray level of the vitreous body region, and standard deviation of gray level of the vitreous body region.
[0019] More preferably, the acquisition of the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis muscle region, and relative gray level standard deviation of the rectus abdominis muscle region using vitreous features as an internal reference benchmark is achieved through the following method: Relative maximum thickness ;in, For maximum thickness, It is the front-to-back diameter; relative average thickness ;in, Average thickness; relative cross-sectional area ;in, For cross-sectional area, The area of the ellipse; Relative major axis length ;in, The length of the major axis. The transverse diameter; Relative minor axis length ;in, The length of the minor axis; The relative average gray level of the rectus abdominis region ;in, The average gray level of the rectus abdominis muscle region. The average gray level of the vitreous region; Relative gray standard deviation of the rectus abdominis muscle region ;in, The standard deviation of grayscale values in the rectus abdominis muscle region. The standard deviation of grayscale in the vitreous region; The relative area, relative volume, relative average gray level, and relative gray level standard deviation of the orbital fat region, obtained using vitreous features as an internal reference benchmark, are obtained through the following methods: relative area ;in, For area; relative volume ;in, For volume; Relative average gray level of orbital fat region ;in, The average gray level of the orbital fat region; Relative gray standard deviation of orbital fat region ;in, The grayscale standard deviation of the orbital fat region; The relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features, obtained using vitreous features as an internal reference benchmark, are obtained through the following method: relative average diameter ;in, The average diameter; Relative maximum diameter ;in, The maximum diameter; Relative minimum diameter ;in, Minimum diameter; Relative optic nerve sheath thickness ;in, The thickness of the nerve sheath; Relative average gray level ;in, This represents the average gray level of the optic nerve region.
[0020] More preferably, the joint feature vector includes: The relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus muscle region, relative gray level standard deviation of the rectus muscle region, roundness, major-minor axis ratio, and ratio of maximum thickness to average thickness of the rectus muscle. The relative area, relative volume, relative average gray level of the orbital fat region, relative gray level standard deviation, contrast, energy, entropy, and homogeneity of the orbital fat region; The relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve region are all characteristics of the optic nerve.
[0021] More preferably, the joint feature vector further includes interaction features, which include: Rectus oculomotor-optic nerve coupling index ; Orbital fat-rectus muscle balance index ; Multi-structure grayscale difference index ; in, This is a smoothing term.
[0022] More preferably, the auxiliary diagnostic results of the thyroid-related ophthalmopathy include: Risk probability value for thyroid-associated ophthalmopathy; Binary classification conclusions derived from risk probability values; In addition, it includes several normalized features and anomaly direction information that contribute the most to the binary classification conclusion.
[0023] Based on the above technical solution, this invention uses a two-dimensional grayscale image of ocular high-frequency linear array ultrasound as input. Through the overall technical chain of "simultaneous segmentation of four structures, extraction of multiple features, normalization with the patient's vitreous body as an internal reference, construction of joint feature vectors, and output of interpretable auxiliary diagnostic results", it achieves interpretable quantification of multiple orbital structures in thyroid-related ophthalmopathy with less imaging and computing resources without requiring three-dimensional reconstruction as a necessary prerequisite.
[0024] This technology chain uses techniques such as vitreous internal reference normalization to ensure quantification capability and comparability. It is this normalization that enables low-cost, easily repeatable ultrasound image data to enter a unified joint feature classification chain. Otherwise, inconsistencies in data across devices and across time will amplify errors.
[0025] Therefore, given that ultrasound examination costs are significantly lower than MRI, this invention has a dual effect: on the one hand, the low examination cost and low equipment threshold reduce the implementation cost of obtaining interpretable multistructural quantification results, enabling outpatient, bedside, and primary care institutions to conduct quantification and follow-up of thyroid-related ophthalmopathy at an affordable cost.
[0026] On the other hand, it improves the efficiency of the stratified use of high-cost imaging resources, freeing MRI from routine multi-structure quantification tasks and focusing it on scenarios requiring three-dimensional reconstruction, assessment of complex orbital lesions, or surgical planning, forming a stratified utilization pattern of "routine ultrasound quantification and targeted MRI review".
[0027] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention, based on ocular ultrasound images, simultaneously quantifies and analyzes the three key pathological structures of thyroid-associated ophthalmopathy: the rectus muscle, orbital fat, and optic nerve. It then constructs a joint feature vector, overcoming the limitations of single-index models. This allows for a more comprehensive capture of imaging changes in different subtypes of thyroid-associated ophthalmopathy, improving the comprehensiveness and accuracy of diagnosis and treatment, and effectively reducing the rate of missed diagnoses.
[0028] 2. This invention utilizes the inherent, relatively stable vitreous features within the eyeball as an internal reference to normalize the ratios of extracted size features, mean grayscale features, and standard deviation grayscale features, while directly preserving dimensionless morphological features and texture statistics. This effectively offsets the effects of differences in gain settings between different ultrasound devices, differences in operation between different operators, and differences in eyeball size between individuals, significantly improving the cross-device and cross-center repeatability and comparability of features.
[0029] 3. This invention uses ocular ultrasound images as the imaging basis, avoiding the ionizing radiation of CT examinations. This makes the method very suitable for outpatient screening, postoperative follow-up, and patients who need multiple re-examinations, which is in line with the trend of clinical convenience and harmlessness.
[0030] 4. The auxiliary diagnostic results output by this invention are based on quantifiable multi-structure imaging features. It can simultaneously output the key features that lead to the judgment of abnormalities and their direction of change, providing intuitive and reliable decision support for clinical practice, and facilitating mutual verification with clinical symptoms and other examination results. Attached Figure Description
[0031] Figure 1 This is a block diagram of the lightweight auxiliary diagnosis and treatment system for thyroid-related ophthalmopathy of the present invention; Figure 2 This is the ROC curve of the model of this invention on an independent test set; Figure 3 This is a statistical graph of the confusion matrix of the independent test set of the present invention; Figure 4 This is a probability calibration curve of the model of the present invention on an independent test set; Figure 5 A statistical chart showing the development set and independent test set samples of this invention; Figure 6 This is a statistical graph showing the performance metrics of the model of this invention on an independent test set; Figure 7 This is a graph showing the predicted probability distribution of the model of this invention on an independent test set; Figure 8 This is a graph showing the top ten permutation importance metrics of the model of this invention on the independent test set; Figure 9 This is a schematic diagram of the multi-anatomical structure segmentation results of the present invention, wherein a is a combined segmentation result of the vitreous body, orbital fat and optic nerve, and b is a segmentation result of the rectus abdominis muscle. Detailed Implementation
[0032] 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.
[0033] It should be noted that the long-standing limitations of ocular ultrasound imaging in the diagnosis and treatment of thyroid-related ophthalmopathy stem from the inherent defects of ultrasound imaging itself: Firstly, ultrasound images suffer from problems such as speckle noise, low tissue contrast, and blurred boundaries, making it difficult to reliably distinguish different structures within the orbit based solely on image grayscale.
[0034] Secondly, the grayscale value of ultrasound images is significantly affected by equipment gain, scanning depth, time gain compensation (TGC) curve, and operator technique. There is a lack of a unified grayscale benchmark among similar structural images acquired by different equipment and different operators.
[0035] Third, ultrasound images are two-dimensional reflection imaging, and the image content is determined by the probe posture, the selection of the section, and the scanning angle. They do not have the spatial correspondence of continuous layers under the unified patient coordinate system of MRI.
[0036] Because of the aforementioned deficiencies, current guidelines use MRI as the gold standard for quantifying orbital structures and assessing thyroid-associated ophthalmopathy. Ultrasound images are only used for the qualitative and initial screening of general intraocular diseases. Current guidelines do not provide any insights into using ocular ultrasound images for quantitative analysis of multiple orbital structures in thyroid-associated ophthalmopathy.
[0037] This application addresses each inherent limitation of ultrasound images by thinking in reverse and successfully constructing corresponding technical means to realize a lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ocular ultrasound images.
[0038] To address speckle noise and low tissue contrast, this application employs noise reduction, CLAHE contrast enhancement, and pixel-physical scale unification sequentially in the preprocessing stage, enabling the stable extraction of the orbital structure boundaries.
[0039] To address the shortcomings of grayscale values and size features being affected by equipment gain, scanning depth, operator technique, and individual eye size, this application uses vitreous features from the same subject and the same scan as an internal reference. It performs class-by-class ratio normalization on the size and grayscale features of the rectus muscle, orbital fat, and optic nerve, so that the non-pathological effects that jointly affect the image are proportionally canceled out. This allows for the acquisition of comparable quantitative features across devices and time without changing the ultrasound acquisition conditions.
[0040] To address the limitation of two-dimensional reflection imaging lacking continuous spatial relationships, this application extracts geometric, grayscale, and texture features with clear anatomical meanings, and constructs joint feature vectors to complete the auxiliary classification of thyroid-related ophthalmopathy.
[0041] Example 1: Lightweight Assisted Diagnosis and Treatment System for Thyroid-Associated Ophthalmopathy Based on Ultrasound Images like Figure 1 As shown, this embodiment provides a lightweight auxiliary diagnosis and treatment system for thyroid-related ophthalmopathy based on ultrasound images, including the following modules: Input module 101 is used to acquire and input the patient's ocular ultrasound images; The region segmentation module 102 is used to segment the rectus abdominis region, orbital fat region, optic nerve region and vitreous region based on ocular ultrasound images; The feature extraction module 103 is used to extract the features of the rectus abdominis muscle in the rectus abdominis muscle region, the features of the orbital fat in the orbital fat region, the features of the optic nerve in the optic nerve region, and the features of the vitreous body in the vitreous body region, respectively. Normalization module 104 is used to normalize the features of the oculomotor muscle, orbital fat and optic nerve respectively, using the vitreous features as the internal reference benchmark, to obtain normalized features of multiple orbital structures. Feature vector module 105 is used to construct a joint feature vector based on normalized features of multiple orbital structures; The classification and diagnosis module 106, based on joint feature vectors, obtains and outputs auxiliary diagnosis and treatment results for thyroid-related eye diseases through a classifier model. The normalization module 104 includes: The rectus abdominis feature quantization unit is used to obtain the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis region, and relative gray level standard deviation of the rectus abdominis region, using the vitreous features as an internal reference benchmark. The orbital fat feature quantification unit is used to obtain the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region with vitreous features as an internal reference. The optic nerve feature quantification unit is used to obtain the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of optic nerve features using vitreous features as an internal reference.
[0042] The rectus abdominis muscle feature quantification unit is used to obtain the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis muscle region, and relative gray level standard deviation of the rectus abdominis muscle region, using vitreous features as an internal reference benchmark. Specifically, this is achieved through the following methods: Relative maximum thickness ;in, For maximum thickness, The anteroposterior diameter is a characteristic feature of the vitreous body; relative average thickness ;in, Average thickness; relative cross-sectional area ;in, For cross-sectional area, The area of the ellipse characteristic of the vitreous body; Relative major axis length ;in, The length of the major axis. The transverse diameter is a characteristic feature of the vitreous body; Relative minor axis length ;in, The length of the minor axis; The relative average gray level of the rectus abdominis region ;in, The average gray level of the rectus abdominis muscle region. The average gray level of the vitreous region; Relative gray standard deviation of the rectus abdominis muscle region ;in, The standard deviation of grayscale values in the rectus abdominis muscle region. The standard deviation of grayscale in the vitreous region; The orbital fat feature quantification unit is used to obtain the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region using vitreous features as an internal reference benchmark. Specifically, it is obtained through the following methods: relative area ;in, For area; relative volume ;in, For volume; Relative average gray level of orbital fat region ;in, The average gray level of the orbital fat region; Relative gray standard deviation of orbital fat region ;in, The grayscale standard deviation of the orbital fat region; The optic nerve feature quantification unit is used to obtain the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features using vitreous features as an internal reference benchmark. Specifically, this is achieved through the following methods: relative average diameter ;in, The average diameter; Relative maximum diameter ;in, The maximum diameter; Relative minimum diameter ;in, Minimum diameter; Relative optic nerve sheath thickness ;in, The thickness of the nerve sheath; Relative average gray level ;in, This represents the average gray level of the optic nerve region.
[0043] Example 2: A lightweight auxiliary diagnostic and treatment method for thyroid-related ophthalmopathy based on ultrasound imaging This embodiment provides a lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images, including the following steps: S1: Acquire ultrasound images of the patient's eye; This step aims to obtain high-quality, repeatable ocular ultrasound images, which form the basis for subsequent analysis. The specific implementation is as follows: 1. Data acquisition equipment A high-frequency linear array probe with a center frequency of 10MHz to 18MHz, preferably 12-15MHz, is used. A two-dimensional grayscale imaging mode is set, with the mechanical index (MI) controlled below 0.23 and the thermal index (TI) controlled below 0.7 to ensure inspection safety. The single-frame image matrix is no less than 512×512, and the grayscale bit depth is 8-bit or 12-bit.
[0044] 2. Equipment parameters The acquisition equipment parameters were set as follows: total gain 55%-70%, dynamic range 45-65dB, focus placed on the depth level from the anterior edge of the optic nerve to the extraocular muscles, and detection depth set to 35mm-50mm. The time gain compensation (TGC) curve should be kept flat or its setting recorded for subsequent reproduction.
[0045] 3. Patient preparation The patient lies supine with eyes gently closed, maintaining eyeballs in their normal position and looking straight ahead. A sterile coupling gel is applied to the skin of the upper eyelid. The probe is then gently placed on the surface of the eyeball through the eyelid to perform a scan, avoiding applying pressure to the eyeball.
[0046] 4. Scanning Specifications The following two basic cross-sections must be collected, and additional oblique cross-sections may be added as needed: Axial section: The probe is placed horizontally in the center of the eyelid to display the transverse sections of the lens, vitreous body, optic nerve, and medial and lateral rectus muscles.
[0047] Sagittal section: The probe is placed vertically in the center of the eyelid. This section should clearly show the long axis of the superior or inferior rectus muscle.
[0048] Nasal or temporal oblique sections: used to complete the boundaries of the medial or lateral rectus muscles.
[0049] 5. Data storage Create an individual file for each patient, storing the original images of all their sections and the corresponding metadata. Following the above guidelines, continuously acquire 3-5 frames of dynamic images for each section of the patient while the patient holds their breath or breathes calmly.
[0050] Select the two images with the clearest boundaries and the fewest artifacts, save them in DICOM or lossless PNG / BMP format, and record all scanning parameters (probe frequency, gain, depth, dynamic range, TGC, acquisition time, eye type, and section type, etc.) metadata.
[0051] 6. Preprocessing of ocular ultrasound images The acquired raw ultrasound images of the eye suffer from speckle noise and low contrast. This step aims to improve image quality and highlight the boundary information of the target structure. The image preprocessing workflow includes denoising, contrast enhancement, and scale unification.
[0052] A1. Speckle Reduction and Noise Reduction Let the original ocular ultrasound image be The expression for midpoint filtering is: ; in, Represented by pixels The center is the filtering window. This represents the ultrasound image after median filtering, where Median represents the median filtering operator. This represents the coordinates of a temporary pixel.
[0053] In this step , , and These represent the image matrix or pixel intensity during the preprocessing stage, respectively; the following text includes structure subscripts. , , and These represent the average gray values of the rectus abdominis muscle region, the orbital fat region, the optic nerve region, and the vitreous body region, respectively.
[0054] A2, Contrast Enhancement The contrast-limited adaptive histogram equalization (CLAHE) algorithm is used to enhance the local contrast of the denoised ocular ultrasound image. CLAHE divides the ocular ultrasound image into several non-overlapping local regions, with a preferred Clip Limit of 2.0-3.0, such as 2.5.
[0055] A3, Brightness Compensation If the preprocessed ocular ultrasound image is still too dark overall, gamma correction can be applied to adjust the global brightness, i.e.: ; The value of gamma_gam is 0.8-1.2, preferably 0.9; when gamma_gam<1, it enhances shadow detail.
[0056] A4. Standardization To ensure the physical comparability of subsequently extracted size features, pixel coordinates need to be converted to physical dimensions. If the image contains a DICOM header, the Pixel Spacing tag is read directly; otherwise, calibration is performed based on the electronic scale embedded in the image, and the scale factor is calculated. (mm / pixel).
[0057] S2: Based on ocular ultrasound images, the rectus abdominis region, orbital fat region, optic nerve region, and vitreous region are segmented; This step is a prerequisite for feature extraction and requires accurate segmentation of the vitreous body, rectus abdominis muscle, orbital fat, and optic nerve. A three-stage strategy of "coarse localization -> fine segmentation -> post-processing" is adopted.
[0058] S201: Vitreous region segmentation; The vitreous body appears as a uniform hypoechoic dark area in ocular ultrasound images, serving as a stable internal reference. In axial sections, coarse localization can be applied to the preprocessed ocular ultrasound image using global thresholding. The threshold can be set to 30-50 of the overall grayscale range (0-255) of the ocular ultrasound image, preferably 40, meaning pixels with a grayscale value less than 40 are considered candidate hypoechoic regions, resulting in a binary mask.
[0059] Fine segmentation and fitting: Morphological closing operations are performed on the binary mask to fill small holes, and the connected regions with the largest areas are extracted as the vitreous body regions. An ellipse was fitted to this region to obtain the anterior-posterior diameter of the vitreous body. transverse diameter And calculate its ellipse area. and the average gray level within the region .
[0060] Simultaneously, the standard deviation of pixel grayscale within the glass region is calculated as an internal reference benchmark for subsequent grayscale standard deviation normalization.
[0061] S202: Rectus oculi region segmentation; Take the segmentation of the medial rectus / lateral rectus muscle (axial view) or the superior rectus / inferior rectus muscle (sagittal view) as an example.
[0062] ROI delineation: The area where the rectus abdominis muscle may be located can be roughly outlined based on prior knowledge (such as its location on both sides of the optic nerve) or by using a lightweight detection network.
[0063] Fine-grained segmentation: Within the Region of Interest (ROI), an active contour model (ACM, Snake model) is used for segmentation. The initial contour is obtained by shrinking the ROI's bounding rectangle by 10%-15%. The energy function is defined as the sum of the internal and external energies. The internal energy controls the smoothness of the contour, while the external energy is defined as the negative value of the image gradient, driving the contour to move towards strong edges. The contour point positions are iteratively updated by minimizing the total energy.
[0064] Active contour model parameter settings: The number of iterations is preferably 200-300, and the smoothness coefficient is... The preferred value is 0.2, and the boundary term weight is... The preferred value is 0.8.
[0065] U-Net can also be used for pixel-level segmentation. The network input is a 256×256 ROI sub-image. The number of encoder channels can be set to 32-64-128-256-512. The loss function is a weighted sum of DiceLoss and BCE, with the preferred weights being 0.7:0.3. The output is a binary mask.
[0066] S203: Segmentation of orbital fat region; Orbital fat is located behind the eyeball, deep to the vitreous body and rectus muscle, and has uneven echo intensity.
[0067] Region definition: Based on the segmented vitreous and rectus abdominis regions, the posterior orbital region is defined as a candidate region for orbital fat in the image.
[0068] Segmentation Implementation: The orbital fat is located behind the eyeball, deep to the vitreous body and the rectus abdominis muscle, and has uneven echo intensity. The posterior orbital region can be defined as a candidate region based on the location of the already segmented vitreous and rectus abdominis muscle regions, and segmented using region growing or graph cut methods.
[0069] When using graph cut, the energy function can be expressed as: ; Among them, boundary weights A value of 20-50 is acceptable, with 30 being the preferred value. Let L represent the energy function, and L represent the set of variables to be optimized. Represents a data item. It represents the paired potential function.
[0070] S204: Localization and segmentation of the optic nerve region; In axial section, the optic nerve appears as a cord-like hypoechoic structure located directly behind the eyeball.
[0071] Location: Search for low-echo elliptical structures with a specific aspect ratio within a depth range of 2-6 mm behind the sphere. Candidate ellipses can be detected using Hough circle / elliptic transform, and then corrected using local gradients and morphological constraints.
[0072] Segmentation: Starting from the located optic nerve center, a region growing algorithm is used, with the gray-level tolerance set to ±15 of the average gray level of the seed point. The optic nerve region is then grown. Its equivalent diameter and the thickness of the optic nerve sheath can be further calculated. If a semi-automatic method is used, the operator clicks on the optic nerve center point, and contour fitting is completed within an 8-12mm search window.
[0073] S205: Post-segmentation processing; Morphological operations (such as opening and closing operations, small connected region removal, and hole filling) are performed on all segmentation results to smooth the boundaries. The removal threshold can be set to pseudo-small blocks smaller than 3%-5% of the area of the target main connected region. If the area difference of corresponding structures in two consecutive frames exceeds 20%, manual review or re-segmentation is triggered.
[0074] S3: Extract the features of the rectus abdominis muscle in the rectus abdominis muscle region, the features of the orbital fat in the orbital fat region, the features of the optic nerve in the optic nerve region, and the features of the vitreous body in the vitreous body region, respectively. This is the core step of this embodiment, which aims to extract four types of features—size, shape, grayscale, and texture—from the segmented image region, perform dimensionless normalization processing using the vitreous body features as an internal reference benchmark, and finally construct a joint feature vector for classification.
[0075] In this embodiment, grayscale features include the region average grayscale and the region grayscale standard deviation; texture statistics are statistical features calculated by methods such as grayscale co-occurrence matrix, and are used separately from the aforementioned grayscale features.
[0076] S301, Characteristics of the oculomotor rectus muscle; The characteristics of the rectus abdominis muscle are 10 dimensions, among which the maximum thickness is directly measured. Average thickness Cross-sectional area Major axis length minor axis length Circularity Major-minor axis ratio Average gray level of the rectus abdominis muscle region Standard deviation of gray values in the rectus abdominis region Derivatives include the ratio of maximum thickness to average thickness. .
[0077] Maximum thickness Defined as the maximum width along the mask normal direction; average thickness Defined as the arithmetic mean of the widths measured column by column along the principal axis; cross-sectional area The product of the number of mask pixels and the square of the scale factor; major axis length. and minor axis length Obtained by principal component analysis or minimum bounding ellipse.
[0078] Circularity Defined as: ; in, Indicates the area of the rectus abdominis muscle covering the area. This indicates the circumference of the rectus abdominis membrane.
[0079] Major-minor axis ratio Defined as: ; in, Indicates the length of the major axis. Indicates the length of the minor axis.
[0080] Maximum thickness to average thickness ratio Defined as: ; The ratio of maximum thickness to average thickness can be used to characterize the degree of focal hypertrophy, but it is not a new feature; rather, it is the 10th dimension of the 10-dimensional features of the rectus abdominis muscle.
[0081] S302, orbital fat characteristics; The orbital fat features are 8-dimensional, including area. ,volume Average gray level of the orbital fat region Standard deviation of grayscale in the orbital fat region Contrast, Energy, Entropy, and Homogeneity.
[0082] area Obtained by multiplying the area of the orbital fat mask by the square of the scale factor; volume When there is multi-faceted data, the area integration method of adjacent facets is used for calculation. In single-faceted mode, the equivalent volume is formed by multiplying the area by the front and back diameters of the glass body.
[0083] In this invention, the volume of orbital fat is referred to as a characteristic feature. This includes the integral volume under multi-section data and the equivalent volume under single-section data; when continuous multi-section data is not available, the equivalent volume is used to characterize the orbital fat size.
[0084] S303, Optic nerve characteristics; The optic nerve features are 5-dimensional, including average diameter. Maximum diameter , minimum diameter optic nerve sheath thickness and the average gray level of the optic nerve region .
[0085] average diameter The calculation method is as follows: ; in, This indicates the area of the optic nerve region mask. Pi is the mathematical constant of a circle.
[0086] Optic nerve sheath thickness The calculation method is as follows: ; in, Indicates the outer diameter of the optic nerve sheath. This indicates the diameter of the optic nerve parenchyma.
[0087] S304, Vitreous characteristics; The vitreous body features are 5-dimensional, including anteroposterior diameter. transverse diameter ellipse area Average gray level of the vitreous region and the standard deviation of grayscale in the vitreous region Normalized outputs are all dimensionless, facilitating comparisons across devices, centers, and patients.
[0088] In this embodiment, the vitreous features mainly serve as an internal reference scale and normalization mapping, and are not used as an independent classification dimension in the final 26-dimensional joint feature vector.
[0089] S4: Using the vitreous features as an internal reference, the features of the oculomotor rectus muscle, orbital fat and optic nerve are normalized to obtain normalized features of multiple orbital structures. Dimensional ratio normalization based on vitreous body features as an internal reference benchmark is key to eliminating differences between equipment and individual components. To avoid the open-ended term "corresponding size features," this invention explicitly defines whether each dimension in the 26-dimensional joint feature vector is normalized and the corresponding vitreous body reference value.
[0090] Symbol explanation: m represents the rectus abdominis muscle, f represents orbital fat, v represents vitreous body; T represents thickness, D represents diameter, A represents area, V represents volume, I represents average gray level, SD represents standard deviation of gray level, and ε is a smoothing term to prevent the denominator from being zero.
[0091] optic nerve-related primitive features were adopted , , , and The suffix _n for each feature parameter symbol indicates only the relative feature after normalization to the internal reference standard of the vitreous body, and does not represent the optic nerve structure.
[0092] It should be noted that this embodiment does not normalize all features in the 26-dimensional joint feature vector, but rather normalizes the ratios of the 16-dimensional basic features that have dimensions or are greatly affected by device grayscale based on the vitreous features.
[0093] For roundness Major-minor axis ratio Ratio of maximum thickness to average thickness Dimensionless morphological features, orbital fat texture statistics, and 3D interactive features constructed from normalized basic features are directly incorporated into the joint feature vector.
[0094] Therefore, the number of normalization formulas and the total number of 26-dimensional joint features do not need to be equal one-to-one.
[0095] S401: For the features of the oculomotor rectus muscle, the normalized result is: Relative maximum thickness ; relative average thickness ; relative cross-sectional area ; Relative major axis length ; Relative minor axis length ; Relative average gray level ; Relative gray standard deviation ; Circularity Major-minor axis ratio and the ratio of maximum thickness to average thickness Since it is a dimensionless morphological feature, it is directly preserved without further normalization.
[0096] S402: For orbital fat features, the normalized result is: relative area ; relative volume ; Relative average gray level ; Relative gray standard deviation ; Contrast, Energy, Entropy, and Homogeneity are texture statistics of the gray-level co-occurrence matrix (GLCM), and are preferably directly incorporated into the joint vector.
[0097] S403: For optic nerve features, the normalization result is: relative average diameter ; Relative maximum diameter ; Relative minimum diameter ; Relative optic nerve sheath thickness ; Relative average gray level .
[0098] Among them, the average diameter of optic nerve features Maximum diameter , minimum diameter and the thickness of the optic nerve sheath All use the transverse diameter of the glass body As an internal reference benchmark.
[0099] Interactive features are constructed using normalized base features, including the oculomotor-optic nerve coupling index, the orbital fat-oculomotor balance index, and the multi-structure grayscale difference index.
[0100] Among them, the rectus abdominis-optic nerve coupling index Using the relative average gray level of the orbital fat region as an intermediate reference, this study characterizes the degree of comprehensive gray level deviation of the rectus muscle and optic nerve relative to the orbital fat, reflecting the synergistic pathological changes among different orbital structures. In this context, 'n' represents the optic nerve identifier in interactive feature naming and is not part of the feature parameter symbol suffix '_n'.
[0101] Orbital fat-rectus muscle balance index , To smooth out the term and prevent the denominator from being a minimum of zero, it is used to distinguish between fat-dominated and muscle-dominated changes.
[0102] Multi-structure grayscale difference index Using the relative average gray level of the orbital fat region as an intermediate reference, this study describes the degree of echo heterogeneity among the rectus muscle, orbital fat, and optic nerve, reflecting the distribution differences of thyroid-associated ophthalmopathy lesions in multiple orbital anatomical structures.
[0103] S5: Construct a joint feature vector based on normalized features of multiple orbital structures; The normalized features of the oculomotor muscle, orbital fat, optic nerve, and interaction features are concatenated into a 26-dimensional joint feature vector in a preset order. Specifically, the features include 10 dimensions of rectus abdominis muscle characteristics, 8 dimensions of orbital fat characteristics, 5 dimensions of optic nerve characteristics, and 3 dimensions of interaction characteristics. Their composition, calculation methods, and normalization rules are shown in Table 1 below: .
[0104] S6: Based on the joint feature vector, the auxiliary diagnosis and treatment results of thyroid-related ophthalmopathy are obtained through a classifier model; S601: Model training, validation, and evaluation; This step aims to build and optimize the machine learning model used for final classification. This invention employs a supervised learning binary classification model to distinguish between thyroid-associated ophthalmopathy and normal samples. The gold standard is derived from clinical diagnosis and treatment results combined with existing imaging findings, including CT or MRI imaging conclusions, independently annotated by at least two ophthalmologists experienced in orbital diseases; if the opinions of two doctors differ, a third senior physician will make the final decision.
[0105] 1. Data preparation and segmentation Ocular ultrasound image data were collected from subjects with clinically diagnosed thyroid-associated ophthalmopathy and healthy volunteers. The dataset was randomly divided into a training set (70%), a validation set (15%), and a test set (15%), based on the subject, to ensure that all images of the same subject appeared in only one set and to prevent data leakage. When the sample size was small, 5-fold cross-validation was used, and an independent test set was reserved in the outer layer.
[0106] 2. Feature Engineering Calculate the mean and standard deviation of each feature on the training set, and perform Z-score standardization on all features in the training, validation, and test sets. Using the training set data, calculate the Pearson correlation coefficient between features; if the absolute value of the correlation coefficient between a pair of features is greater than 0.9, one of them is removed. Alternatively, recursive feature elimination combined with cross-validation can be used to select the optimal feature subset.
[0107] 3. Training the classifier Support Vector Machines (SVMs) are preferred. For linearly inseparable scenarios, a Radial Basis Function (RBF) kernel is used. The kernel function parameters γ and the penalty coefficient C are optimized on the validation set using a grid search. The preferred search range is C∈{0.1,1,10,50,100}, γ∈{0.001,0.01,0.1,1}. When the data dimension is low and close to linearly separable, a linear kernel can also be used as an alternative.
[0108] 4. Model Evaluation Evaluate the performance of the final model on independent test sets, and report the area under the receiver operating characteristic (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, and calibration metrics (such as the Brier score). Also plot the ROC curve, confusion matrix, and calibration curve.
[0109] 5. Evaluation of Multi-Structure Segmentation Modules
[0110] For the multi-structure segmentation module itself, it is preferable to use the manually drawn results as a reference, and report the Dice coefficient, intersection-over-union ratio (IoU), Hausdorff distance, and average surface distance. Preferably, the Dice values for the rectus muscle are ≥0.85, the vitreous body Dice are ≥0.90, and the optic nerve Dice are ≥0.80; if they are below the above thresholds, it indicates that the segmentation results are not reliable enough, and it is recommended to retrain or optimize the segmentation model.
[0111] 6. Robustness and repeatability assessment For the same subject, at least two ultrasound scans should be performed in the same examination, and the intraclass correlation coefficient (ICC) or coefficient of variation (CV) of the main features should be calculated. Core dimensional features should have an ICC ≥ 0.80 or a CV ≤ 10%. When validating across devices, at least two different models or with different gain settings of ultrasound equipment should be compared.
[0112] In a reproducible embodiment, clinically diagnosed thyroid-associated ophthalmopathy patients and age- and sex-matched normal controls can be collected and randomly divided into training, validation, and test sets according to the patient level. All raw images undergo uniform preprocessing, and two doctors delineate the vitreous body, target rectus muscle, orbital fat, and optic nerve to establish segmentation training and evaluation benchmarks.
[0113] The segmentation network uses 2DU-Net with an input size of 256×256, a batch size of 8, and an initial learning rate of 1×10⁻⁶. -3 The optimizer is Adam, with a weight decay of 1×10⁻⁶. -5 The training run consists of 100-150 epochs, and the test is stopped early if there is no improvement after 15 consecutive epochs on the validation set. Data augmentation includes horizontal flipping, ±10° rotation, 0.9-1.1x scaling, brightness perturbation, and mild Gaussian noise.
[0114] In the classification phase, RBF-SVM is used on the extracted 26-dimensional joint feature vector, and five-fold cross-validation is employed for hyperparameter search. Finally, ROC, AUC, sensitivity, and specificity metrics are output on the independent test set. To improve model interpretability, SHAP can be used to rank feature contributions.
[0115] The method of this invention can simultaneously and clearly delineate the boundaries of four types of anatomical structures—vitreous body, retro-orbital tissue, optic nerve, and rectus abdominis muscle—in a single frame of two-dimensional grayscale ultrasound image. The segmentation quality is reliable, laying a solid foundation for feature extraction.
[0116] Based on the anatomical structure segmentation results, the RBF-SVM classifier was used to classify the 26-dimensional joint feature vector, and good classification performance was achieved on the independent test set.
[0117] Figure 2 The ROC curves for the independent test set are shown, with an area under the curve (AUC) of 0.833 and a Bootstrap 95% confidence interval of 0.705–0.935, indicating that the model has good discriminative ability.
[0118] Figure 3 The confusion matrix for the independent test set (71 eyes) shows 25 true negatives, 7 false positives, 7 false negatives, and 32 true positives. Based on this, the sensitivity is 82.1%, the specificity is 78.1%, the positive predictive value is 82.1%, and the negative predictive value is 78.1%.
[0119] Figure 4 The model's probability calibration curve on the independent test set is used to evaluate the consistency between the predicted probability and the actual frequency of occurrence by comparing the current model calibration curve with the ideal calibration line, and a Brier score of 0.177 is given.
[0120] Figure 5 The data composition of the development set and the independent test set is shown. The development set contains 75 subjects and 170 eyes that passed quality control, while the independent test set contains 32 subjects and 71 eyes that passed quality control. The number of subjects and the number of eyes that passed quality control are shown in the figure as bar charts.
[0121] Figure 6 The model's performance metrics on the independent test set are shown, including AUC 83.3%, accuracy 80.3%, sensitivity 82.1%, specificity 78.1%, positive predictive value (PPV) 82.1%, negative predictive value (NPV) 78.1%, and F1 score 82.1%, and each metric is presented visually in a bar chart.
[0122] Figure 7 The predicted probability distribution of the model on the independent test set is given. The normal group has 32 eyes and the thyroid-associated ophthalmopathy group has 39 eyes. The dashed line represents the decision threshold of 0.5.
[0123] Figure 8 The machine error range of the top ten indicators of permutation importance of the model on the independent test set is given. The horizontal axis represents the decrease in AUC after permutation. The larger the value, the greater the contribution of the corresponding indicator to the model's judgment.
[0124] S602: Auxiliary diagnostic and treatment output and quality control; The normalized joint feature vector obtained after processing S1-S5 is input into the classification model trained in step S601 to obtain the auxiliary diagnosis and treatment results of thyroid-related ophthalmopathy.
[0125] Auxiliary diagnostic results should include at least: risk probability value for thyroid-associated ophthalmopathy. The binary classification conclusion is either "Thyroid-associated ophthalmopathy imaging abnormalities" or "No obvious thyroid-associated ophthalmopathy imaging abnormalities are observed"; key features and their relative abnormal directions are analyzed, and the Z-score of each feature relative to the normal mean is calculated. The top three features with the largest absolute Z-scores and their abnormal directions are listed; gray zone case suggestions are provided.
[0126] To ensure clinical usability, the present invention sets the following quality control rules: no auxiliary diagnostic conclusions are given when the number of analyzed frames is less than 2; when vitreous segmentation fails, the optic nerve boundary is unclear, or more than 20% of the core features are missing, the system automatically prompts "Insufficient image quality, please re-acquire"; when the input data exceeds the distribution range of the feature in the training set, an out-of-distribution (OOD) warning is triggered, and the system automatically prompts "Some feature values are outside the normal range, please make a comprehensive judgment based on clinical findings".
[0127] Example 3: A Complete Link That Can Be Implemented Taking a single axial image of a patient as an example, the vitreous features are segmented as follows: anteroposterior diameter 21.0 mm, glass body transverse diameter 23.5mm, elliptical area of the glass body The average gray level of the vitreous region is 387.0 mm². The gray standard deviation of the vitreous region is 42.0. It is version 6.0.
[0128] Characteristics of the rectus abdominis muscle: maximum thickness It is 5.8mm thick with an average thickness. 4.9mm, cross-sectional area 58.0 mm², major axis length 14.5mm, minor axis length The average gray level of the rectus abdominis muscle region is 5.2 mm. The gray standard deviation of the rectus abdominis muscle region is 61.0. It is 8.1.
[0129] Orbital fat characteristics: area 96.0 mm², volume The average gray level of the orbital fat region is 2016.0 mm³. The gray standard deviation of the orbital fat region is 56.0. The coefficient of performance is 7.5, the contrast ratio is 2.83, the energy is 0.41, the entropy is 2.08, and the homogeneity is 0.74.
[0130] Optic nerve characteristics: average diameter 4.6mm, maximum diameter 5.0mm minimum diameter The thickness of the optic nerve sheath is 4.1 mm. The average gray level of the optic nerve region is 0.9 mm. It is 48.0.
[0131] According to the dimension-by-dimensional normalization rule in Example 2, the relative maximum thickness of the rectus muscle can be obtained. The relative average thickness is 0.276. 0.233, relative cross-sectional area 0.150, relative to the major axis length 0.617, relative minor axis length The relative average gray level of the rectus abdominis muscle region is 0.221. The relative gray standard deviation of the rectus abdominis muscle region is 1.452. 1.350.
[0132] Relative area of orbital fat 0.248, relative volume The relative average gray level of the orbital fat region is 0.248. The relative gray standard deviation of the orbital fat region is 1.333. It is 1.250.
[0133] The relative average diameter of the optic nerve The relative maximum diameter is 0.196. The relative minimum diameter is 0.213. The relative thickness of the optic nerve sheath is 0.174. The relative average gray level of the optic nerve region is 0.038. It is 1.143.
[0134] Furthermore, the rectus abdominis-optic nerve coupling index was obtained. The orbital fat-rectus muscle balance index is 0.054. The multi-structure grayscale difference index is 1.653. The value is 0.309. The 26-dimensional joint feature vector table is concatenated according to the order in Table 1 and input into the classifier model, which outputs the risk probability of thyroid-associated eye disease, the binary classification conclusion, and the top three key explanatory features.
[0135] The aforementioned 26-dimensional joint feature vector preserves circularity. Major-minor axis ratio The feature incorporates dimensionless morphological and textural information such as texture statistics, and eliminates differences in device gain, operator technique, and individual eye size by using the vitreous body as an internal reference, thus enabling robust comparability of features across devices and centers.
[0136] After inputting the joint feature vector into the classifier model, the system can output the risk probability of thyroid-associated ophthalmopathy, the binary classification conclusion, and the top three key explanatory features, thereby providing comprehensive, robust, and interpretable radiation-free auxiliary diagnostic and treatment evidence for clinical practice.
[0137] like Figure 9 As shown, Figure 9 In the middle section 'a', the result of combined segmentation of the vitreous body, orbital fat, and optic nerve is shown. Figure 9 Figure b shows the segmentation result of the oculomotor rectus muscle. As can be seen from the figure, the method of the present invention can clearly delineate the boundaries of four types of anatomical structures—the vitreous body, orbital fat, optic nerve, and oculomotor rectus muscle—in a two-dimensional grayscale ultrasound image.
[0138] The vitreous region appears as a uniform hypoechoic dark area with a complete shape and smooth boundaries. After threshold segmentation and ellipse fitting, its anteroposterior and transverse diameters are consistent with its anatomical location. The rectus abdominis region appears as a spindle-shaped hypoechoic band with continuous boundaries and clear demarcation from the adipose tissue. The optic nerve region appears as a cord-like structure with accurate positioning and clear direction. The posterior orbital tissue region has complete boundaries and connects naturally with the structures within the eyeball and muscle cone.
[0139] There are no overlapping, adhering, or missing regions among the four types of regions, and the segmentation results can be directly used for the extraction of various geometric, grayscale, and texture features. The segmentation quality ensures that the extracted features can truly reflect the anatomical and pathological state of each structure, and class-by-class normalization is performed using vitreous features as an internal reference, providing a reliable foundation for the construction of joint feature vectors and the auxiliary diagnosis and treatment of thyroid-related ophthalmopathy.
[0140] In summary, this case fully verifies the feasibility and effectiveness of the technical solution of the present invention, showing that the method can simultaneously quantify the three key pathological structures of the oculomotor muscle, orbital fat, and optic nerve, and achieve standardization by using internal vitreous reference, effectively improving the accuracy, repeatability, and clinical applicability of auxiliary diagnosis and treatment of thyroid-related ophthalmopathy.
[0141] Example 4: Preferred Implementation Scheme Based on Example 2, several optional implementation methods are further refined and provided, all of which fall within the protection scope of this invention.
[0142] 1. Replacement of data acquisition equipment The acquisition device of this invention can use an 8-14MHz variable frequency linear array probe, a 20MHz ultrasonic biological microscope, or a three-dimensional matrix array probe. When using a three-dimensional probe, the volume can be calculated directly at the voxel level, eliminating the need for multi-faceted area integration.
[0143] 2. Image source replacement In some embodiments, when the patient has already acquired orbital MRI data, the vitreous intra-reference normalization concept of this invention can be used to perform normalization analysis on the features of the rectus abdominis muscle, orbital fat, and optic nerve in the MRI images. When the patient already has orbital MRI data, T2 or fat-suppressed MRI images can be used as input. In this case, the vitreous body shows a uniform high signal on T2WI, and its signal intensity and eye size can also serve as a stable internal reference for normalizing the features of the rectus abdominis muscle, orbital fat, and optic nerve segmented from the MRI images. The preprocessing steps need to be adjusted accordingly to MRI-related corrections.
[0144] 3. Segmentation algorithm replacement The segmentation of each orbital anatomical structure can be achieved using any advanced medical image segmentation algorithm.
[0145] Vitreous segmentation: cueing segmentation can be performed using U-Net and its variants (such as Attention U-Net), Otsu, nnUNet, or the latest segmentation base models (such as SAM).
[0146] Oculomotor muscle segmentation: A two-stage strategy model can be adopted, which first uses YOLO for detection and then combines it with U-Net for fine segmentation.
[0147] Orbital fat segmentation: GraphCuts, random walk, or Transformer segmentation models can be used.
[0148] Optic nerve segmentation: Template matching, polar coordinate contour fitting, or semi-automatic interactive segmentation can be used.
[0149] 4. Classifier replacement Besides SVM, classification models can employ Random Forest, Gradient Boosting Machine (such as XGBoost, LightGBM), Logistic Regression, Lightweight Fully Connected Neural Networks, or 3D-CNN. As long as the model's input is the aforementioned multi-structure joint feature vector normalized based on vitreous reference, it falls under an equivalent implementation of this invention.
[0150] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0151] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A thyroid-associated ophthalmopathy light-weighted assistant diagnosis and treatment system based on ultrasonic images, characterized in that, Includes the following modules: The input module is used to acquire and input the patient's ocular ultrasound images; The region segmentation module is used to segment the rectus abdominis region, orbital fat region, optic nerve region, and vitreous region based on ocular ultrasound images; The feature extraction module is used to extract features of the rectus abdominis muscle in the rectus abdominis muscle region, features of the orbital fat in the orbital fat region, features of the optic nerve in the optic nerve region, and features of the vitreous body in the vitreous body region, respectively. The normalization module is used to normalize the features of the oculomotor rectus muscle, orbital fat and optic nerve respectively, using the vitreous features as the internal reference benchmark, to obtain normalized features of multiple orbital structures. The feature vector module is used to construct joint feature vectors based on normalized features of multiple orbital structures; The classification and diagnosis module, based on joint feature vectors, obtains and outputs auxiliary diagnosis and treatment results for thyroid-related eye diseases through a classifier model. The normalization module includes: The rectus abdominis feature quantization unit is used to obtain the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis region, and relative gray level standard deviation of the rectus abdominis region, using the vitreous features as an internal reference benchmark. The orbital fat feature quantification unit is used to obtain the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region with vitreous features as an internal reference. The optic nerve feature quantification unit is used to obtain the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of optic nerve features using vitreous features as an internal reference.
2. The ultrasonographic image-based thyroid-associated ophthalmopathy light-weighted assist-diagnosis and treatment system according to claim 1, characterized by, The rectus abdominis muscle feature quantification unit is used to obtain the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus abdominis muscle region, and relative gray level standard deviation of the rectus abdominis muscle region, using vitreous features as an internal reference benchmark. Specifically, this is achieved through the following methods: relative maximum thickness ; wherein, is the maximum thickness, is the anterior-posterior diameter of the vitreous body characteristics; relative average thickness ; wherein, is the average thickness; relative cross-sectional area ;in, For cross-sectional area, The area of the ellipse characteristic of the vitreous body; Relative major axis length ;in, The length of the major axis. The transverse diameter is a characteristic feature of the vitreous body; Relative minor axis length ;in, The length of the minor axis; The relative average gray level of the rectus abdominis region ;in, The average gray level of the rectus abdominis muscle region. The average gray level of the vitreous region; Relative gray standard deviation of the rectus abdominis muscle region ;in, The standard deviation of grayscale values in the rectus abdominis muscle region. denoted as the standard deviation of grayscale in the vitreous region.
3. The lightweight auxiliary diagnostic and treatment system for thyroid-related ophthalmopathy based on ultrasound images according to claim 2, characterized in that, The orbital fat feature quantification unit is used to obtain the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region using vitreous features as an internal reference benchmark. Specifically, it is obtained through the following methods: relative area ;in, For area; relative volume ;in, For volume; Relative average gray level of orbital fat region ;in, The average gray level of the orbital fat region; Relative gray standard deviation of orbital fat region ;in, The gray standard deviation represents the orbital fat region.
4. The lightweight auxiliary diagnostic and treatment system for thyroid-related ophthalmopathy based on ultrasound images according to claim 3, characterized in that, The optic nerve feature quantification unit is used to obtain the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features using vitreous features as an internal reference benchmark. Specifically, this is achieved through the following methods: relative average diameter ;in, The average diameter; Relative maximum diameter ;in, The maximum diameter; Relative minimum diameter ;in, Minimum diameter; Relative optic nerve sheath thickness ;in, The thickness of the nerve sheath; Relative average gray level ;in, This represents the average gray level of the optic nerve region.
5. A lightweight auxiliary diagnostic and treatment method for thyroid-related ophthalmopathy based on ultrasound imaging, characterized in that, Includes the following steps: Acquire ultrasound images of the patient's eyes; Based on ocular ultrasound images, the rectus abdominis region, orbital fat region, optic nerve region, and vitreous region were segmented. Features of the rectus abdominis muscle in the rectus abdominis muscle region, features of the orbital fat in the orbital fat region, features of the optic nerve in the optic nerve region, and features of the vitreous body in the vitreous body region were extracted respectively. Using vitreous features as an internal reference, the features of the rectus abdominis muscle, orbital fat, and optic nerve were normalized to obtain normalized features of multiple orbital structures, including: Using vitreous features as an internal reference, the relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus muscle region, and relative gray level standard deviation of the rectus muscle region are obtained. Using vitreous features as an internal reference, the relative area, relative volume, relative average gray level of the orbital fat region, and relative gray level standard deviation of the orbital fat region were obtained. Using vitreous features as an internal reference, the relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features were obtained. Construct a joint feature vector based on the normalized features of multiple orbital structures; Based on joint feature vectors, the auxiliary diagnosis and treatment results of thyroid-related ophthalmopathy are obtained through a classifier model.
6. The lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images according to claim 5, characterized in that, The features of the rectus abdominis muscle include: maximum thickness, average thickness, cross-sectional area, long axis length, short axis length, average gray level of the rectus abdominis muscle region, standard deviation of gray level of the rectus abdominis muscle region, roundness, ratio of long axis to short axis, and ratio of maximum thickness to average thickness. The orbital fat features include: area, volume, average gray level of the orbital fat region, standard deviation of gray level of the orbital fat region, contrast, energy, entropy, and homogeneity. The optic nerve features include: average diameter, maximum diameter, minimum diameter, optic nerve sheath thickness, and average gray level of the optic nerve region; The characteristics of the vitreous body include: anteroposterior diameter, transverse diameter, elliptical area, average gray level of the vitreous body region, and standard deviation of gray level of the vitreous body region.
7. The lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images according to claim 6, characterized in that: The relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus muscle region, and relative gray level standard deviation of the rectus muscle region, obtained by using vitreous features as an internal reference benchmark, are obtained through the following methods: Relative maximum thickness ;in, For maximum thickness, It is the front-to-back diameter; relative average thickness ;in, Average thickness; relative cross-sectional area ;in, For cross-sectional area, The area of the ellipse; Relative major axis length ;in, The length of the major axis. The transverse diameter; Relative minor axis length ;in, The length of the minor axis; The relative average gray level of the rectus abdominis region ;in, The average gray level of the rectus abdominis muscle region. The average gray level of the vitreous region; Relative gray standard deviation of the rectus abdominis muscle region ;in, The standard deviation of grayscale values in the rectus abdominis muscle region. The standard deviation of grayscale in the vitreous region; The relative area, relative volume, relative average gray level, and relative gray level standard deviation of the orbital fat region, obtained using vitreous features as an internal reference benchmark, are obtained through the following methods: relative area ;in, For area; relative volume ;in, For volume; Relative average gray level of orbital fat region ;in, The average gray level of the orbital fat region; Relative gray standard deviation of orbital fat region ;in, The grayscale standard deviation of the orbital fat region; The relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve features, obtained using vitreous features as an internal reference benchmark, are obtained through the following method: relative average diameter ;in, The average diameter; Relative maximum diameter ;in, The maximum diameter; Relative minimum diameter ;in, Minimum diameter; Relative optic nerve sheath thickness ;in, The thickness of the nerve sheath; Relative average gray level ;in, This represents the average gray level of the optic nerve region.
8. The lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images according to claim 7, characterized in that, The joint feature vector includes: The relative maximum thickness, relative average thickness, relative cross-sectional area, relative major axis length, relative minor axis length, relative average gray level of the rectus muscle region, relative gray level standard deviation of the rectus muscle region, roundness, major-minor axis ratio, and ratio of maximum thickness to average thickness of the rectus muscle. The relative area, relative volume, relative average gray level of the orbital fat region, relative gray level standard deviation, contrast, energy, entropy, and homogeneity of the orbital fat region; The relative average diameter, relative maximum diameter, relative minimum diameter, relative optic nerve sheath thickness, and relative average gray level of the optic nerve region are all characteristics of the optic nerve.
9. The lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images according to claim 8, characterized in that, The joint feature vector also includes interaction features, which include: Rectus oculomotor-optic nerve coupling index ; Orbital fat-rectus muscle balance index ; Multi-structure grayscale difference index ; in, This is a smoothing term.
10. The lightweight auxiliary diagnosis and treatment method for thyroid-related ophthalmopathy based on ultrasound images according to claim 5, characterized in that, The auxiliary diagnostic results for thyroid-related ophthalmopathy include: Risk probability value for thyroid-associated ophthalmopathy; Binary classification conclusions derived from risk probability values; In addition, it includes several normalized features and anomaly direction information that contribute the most to the binary classification conclusion.
Citation Information
Patent Citations
OCT image recommendation method and device based on artificial intelligence, equipment and medium
CN114782337A
Thyroid-related eye disease auxiliary verification method based on SPECT / CT and federated learning
CN120689316A
