A multi-dimensional quantitative assessment method and system combining meibomian gland morphology and function

Abstract

This invention discloses a multi-dimensional quantitative assessment method and system combining meibomian gland morphology and function, belonging to the fields of ophthalmic diagnosis and artificial intelligence technology. It acquires color images of the eyelid margin, infrared images of the meibomian glands, and dynamic videos of gland expulsion upon pressure, inputting these into a dedicated AI analysis engine for processing. Employing an innovative anatomically guided backbone network and a meibomian gland region adaptive module, it achieves parallel automatic quantitative scoring across four dimensions: meibomian gland opening abnormality score, secretion characteristics score, expulsion capacity score, and morphological abnormality score based on gland segmentation loss rate. The four scores are summed to obtain a comprehensive total score, used for the graded diagnosis of meibomian gland dysfunction. By comparing changes in the patient's scores over time, it achieves macroscopic and microscopic evaluation of treatment efficacy. By integrating functional and morphological indicators into a unified automated quantitative system, it solves the problems of subjectivity and fragmentation in existing assessment methods, significantly improving the objectivity, comprehensiveness, and accuracy of diagnosis, and providing a reliable tool for individualized treatment and long-term management of MGD.

Description

Technical Field

[0001] This invention relates to the fields of ophthalmic diagnostics and artificial intelligence technology, and in particular to a multi-dimensional quantitative assessment method and system that combines meibomian gland morphology and function. Background Technology

[0002] Meibomian gland dysfunction (MGD) is a major cause of dry eye, and its pathophysiological changes involve two main aspects: functional changes and structural (morphological) changes. Functional changes include blockage of the meibomian gland openings, abnormal characteristics of secretions, and decreased drainage capacity. Structural changes are mainly manifested as atrophy, absence, dilation, or distortion of the meibomian glands.

[0003] In current clinical practice, the assessment of these two aspects of change is often fragmented. Physicians assess function through slit-lamp observation and palpation, and observe morphology using infrared meibomian gland imaging (Meibography). This separate assessment method lacks a comprehensive, quantitative scoring system that can organically combine the two. Furthermore, the interpretation of meibomian gland morphology in infrared images relies heavily on the physician's subjective experience, resulting in low efficiency and poor consistency.

[0004] Patent CN201911349411A and other technologies propose using AI to automatically segment and quantify meibomian gland morphology, solving the problem of automating morphological assessment. However, these technologies do not involve integration with functional indicators. Therefore, developing a method and system that can automatically and synchronously assess meibomian gland "morphology" and "function" and integrate them into a unified scoring system has significant clinical value for the accurate diagnosis, classification, and treatment guidance of MGD. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-dimensional quantitative assessment method and system that combines meibomian gland morphology and function.

[0006] This invention is achieved through the following technical solution: a multi-dimensional quantitative assessment method combining meibomian gland morphology and function, comprising the following steps:

[0007] Multimodal information acquisition steps: Acquire patient eyelid multimodal data including color static images of the eyelid margin, infrared images of the meibomian glands, and dynamic video of secretions being discharged upon pressure;

[0008] The four-dimensional automatic scoring process involves inputting the multimodal data into an AI analysis model, which then outputs quantitative scores across four dimensions in parallel or sequentially, including:

[0009] a. Meibomian gland orifice abnormality score;

[0010] b. Score of meibomian gland secretion characteristics;

[0011] c. Meibomian gland excretion capacity score;

[0012] d. Meibomian gland absence area score, which is obtained by calculating the absence rate and mapping it based on the gland segmentation results in the infrared image;

[0013] Comprehensive assessment steps: The scores of the four dimensions are added together to obtain the total score of meibomian gland function and morphology, and the meibomian gland dysfunction is graded according to the total score range;

[0014] Treatment efficacy tracking steps: Compare the scores of the four dimensions and the overall score of the same patient at different time points to assess the treatment effect and analyze the improvement dimensions.

[0015] Furthermore, the AI ​​analysis model includes:

[0016] The anatomically guided backbone network has at least some of its convolutional layers using deformable strip convolutions, and the convolution offset is constrained by the eyelid curvature prior map.

[0017] The meibomian gland region adaptive module is used for dynamic contour-aware pooling based on the gland presence probability map and conditional modulation of features based on missing modes.

[0018] The four-dimensional scoring subnetwork is used to perform opening abnormality identification, secretion morphology classification, expulsion event detection, and gland segmentation and missing rate calculation.

[0019] Furthermore, the meibomian gland region adaptive module includes:

[0020] Glandular contour probability map prediction unit;

[0021] Pooling units that generate an adaptive sampling grid based on the probability graph;

[0022] A missing pattern classifier is used to output the classification of glandular absence degree and to conditionally modulate features for subsequent tasks.

[0023] Furthermore, the meibomian gland absence area scoring step includes:

[0024] Use U-Net or a neural network with an improved U-Net architecture to perform gland pixel-level segmentation in infrared images;

[0025] The meibomian gland absence rate is calculated using the formula: (Total palpebral conjunctival area / Total gland area) × 100%.

[0026] Based on the preset missing rate threshold range, the missing rate is mapped to a score of 0-3.

[0027] Furthermore, the secretion phenotype score is obtained based on dynamic video analysis, specifically including:

[0028] Temporal convolutional networks or ConvLSTM are used to extract spatiotemporal features of color, transparency and morphological changes during the secretion process;

[0029] The extracted features are classified as clear, cloudy, granular, or toothpaste-like, and mapped to a score of 0-3.

[0030] The discharge capacity score is obtained based on dynamic video analysis, specifically including:

[0031] The gland openings that successfully expel secretions during the pressing process are identified using optical flow or trajectory tracking technology.

[0032] The proportion of successfully discharged openings to the total number of pressurized openings was statistically analyzed and mapped to a score of 0-3.

[0033] Furthermore, the method also includes:

[0034] During the training phase of the AI ​​analysis model, a multi-task loss function is used for joint optimization. The loss function includes loss terms for each dimension of scoring tasks and a logical consistency loss term, which is used to constrain the pathological rationality between morphological deficiencies and functional scores.

[0035] A multi-dimensional quantitative assessment system combining meibomian gland morphology and function was also disclosed, including:

[0036] A multimodal acquisition device is used to acquire color static images of the eyelid margin, infrared images of the meibomian glands, and dynamic videos of secretion discharge.

[0037] The AI ​​analysis engine is deployed with the AI ​​analysis model as described in claims 2-3, which is used to receive the multimodal data and output a four-dimensional score and a comprehensive total score;

[0038] The report generation unit integrates scoring results, visualizations, and historical comparison information to generate a structured evaluation report.

[0039] Furthermore, the AI ​​analysis engine further includes:

[0040] The data preprocessing module is used for spatial registration, size normalization, image enhancement, and video keyframe extraction of multimodal images;

[0041] The model interpretation and visualization module is used to generate gland segmentation overlay maps, opening anomaly annotation maps, and high-risk area heat maps.

[0042] The beneficial effects of this invention are as follows:

[0043] 1. Using AI to replace subjective human judgment enables the automation, standardization, and quantification of meibomian gland assessment, resulting in more objective and repeatable results and greatly improving examination efficiency.

[0044] 2. For the first time, it quantitatively combines four major indicators: whether the gland is missing in shape, whether its function is blocked, the quality of the secretions, and whether they can be discharged, providing a more comprehensive diagnostic perspective than a single examination.

[0045] 3. Automatically compares patients' historical data, showing not only changes in the total score but also clear details of subtle improvements in each sub-indicator, accurately assessing treatment effectiveness and guiding adjustments to the treatment plan.

[0046] 4. It supports both cloud-based analysis and localized software modes, making it easy to promote and apply in primary hospitals, improving the early screening and diagnosis capabilities of MGD, while ensuring data security.

[0047] 5. Through special design, the AI ​​model's analysis process aligns with medical logic, such as the correlation between glandular absence and functional decline, thereby enhancing the credibility of the results and doctors' acceptance. Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 This is a schematic diagram of the arc-shaped mask for the upper and lower eyelids of the present invention;

[0050] Figure 3 This is a schematic diagram comparing the square receptive field of traditional convolution with the serpentine receptive field of the strip convolution of this invention.

[0051] Figure 4 This is a structural diagram of the TAAG module of the present invention;

[0052] Figure 5 This is a schematic diagram comparing the sampling points of square convolution and strip deformable convolution according to the present invention;

[0053] Figure 6 This is a diagram of the offset prediction network structure of the present invention; Detailed Implementation

[0054] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0055] 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.

[0056] Example 1

[0057] This invention proposes a multi-dimensional quantitative assessment system that combines meibomian gland morphology and function, which is particularly suitable for the automated and comprehensive diagnosis and efficacy evaluation of meibomian gland dysfunction (MGD).

[0058] The multi-dimensional quantitative assessment system combining meibomian gland morphology and function in this embodiment mainly consists of three parts:

[0059] 1. Multimodal image acquisition device: including a digital photographic slit lamp with white light illumination and high-definition video recording, a meibomian gland imager with an integrated infrared imaging module (or a standalone infrared imaging device), and a video recording unit for recording the secretion discharge process under standard compression. These devices can be integrated into a single device or connected to a central processing unit via a network.

[0060] 2. AI Comprehensive Analysis Engine (Server / Cloud Platform): Deployed with a trained multi-task neural network model, responsible for receiving and processing multimodal data and performing automatic scoring. This includes high-performance servers deployed locally in the hospital or in the cloud. Built-in high-performance GPU: Configured with professional-grade graphics processors (such as NVIDIA A100, V100, etc.) to accelerate complex calculations in the neural network; High-speed network: Secure, high-speed local area network or internet connection within the hospital for stable and rapid transmission of image and video data from acquisition devices (slit lamps, infrared imagers) to the analysis engine. Data storage system: Utilizes server or cloud storage to securely store patients' raw multimodal data, AI analysis results, historical reports, and the trained AI model itself.

[0061] The specific processing flow is as follows: First, data reception and preprocessing: The engine receives eyelid margin color images, infrared images, and dynamic videos from the acquisition end via a network interface. Upon receipt, a series of standardized operations are immediately performed, including: Image alignment: Affine transformation is used to spatially align images of different modalities to ensure that the same region is being analyzed. Size unification: All images are scaled to a uniform size (e.g., 256x256 pixels) to meet the model input requirements. Video frame extraction: Keyframes (e.g., seconds 3, 6, 9, and 12) are extracted from the pressure video to form a temporal sequence. Image enhancement: Image quality is improved and lesion features are highlighted through techniques such as color normalization and contrast enhancement. Then, AI feature extraction and scoring are performed. The preprocessed data is fed into the core multi-task neural network for analysis: an improved ResNet-50 network is used to efficiently extract deep features from the images. The system intelligently identifies effective glandular regions, allowing the model to focus its attention on these areas and dynamically adjust subsequent analyses based on the pattern of glandular absence. Different "branches" of the model simultaneously analyze the data, outputting scores across four dimensions (MGO, MGS, MGA, MGL) within seconds. Finally, the results aggregation and report generation engine adds the four scores to obtain a comprehensive total score and determines the severity level of MGD (mild / moderate / severe) based on a preset threshold (e.g., 0-12 points). Furthermore, the engine integrates all quantitative scores, diagnostic grading, and visualizations (e.g., glandular segmentation maps, orifice obstruction annotation maps) into a structured electronic report. The generated report is sent to the doctor's workstation for review. Simultaneously, all analysis results, along with the raw data, are encrypted and archived in the patient's electronic medical record for future efficacy tracking and comparison.

[0062] 3. User interaction and report generation terminals: such as doctors' workstation computers, tablets, etc., used to submit case data, view AI analysis results (including various scores, visualized images, historical comparison curves), and generate structured diagnostic reports.

[0063] Example 2

[0064] like Figure 1 As shown: This embodiment also discloses a multi-dimensional quantitative assessment method combining meibomian gland morphology and function, and the detailed implementation steps are as follows:

[0065] Step S1: Standardized multimodal information acquisition

[0066] 1. Patient preparation and positioning: The patient places their chin on the slit lamp chin rest, with their forehead pressed against the headband, and the operator guides their eyes to focus in the designated direction.

[0067] 2. Functional Image Acquisition: Adjust the slit lamp light source to diffuse white light, focus on the middle section of the lower eyelid margin, and take a color static image of the eyelid margin that clearly shows at least 8-12 consecutive meibomian gland openings (recommended resolution ≥1920×1080 pixels). Ensure that details such as ester caps, bulges, or fat plugs at the openings are clearly distinguishable in the image.

[0068] 3. Morphological Image Acquisition: Rotate the patient's upper and lower eyelids and use infrared imaging mode (wavelength typically 850nm or 940nm) to capture infrared images of the meibomian glands in the upper and lower eyelids respectively. The images should clearly show the glandular direction, branches, and missing areas of the meibomian glands (resolution ≥1280×960 pixels is recommended).

[0069] 4. Dynamic Information Acquisition: Apply standardized, constant pressure (e.g., using a calibrated ophthalmic swab or a dedicated pressure-sensing probe, with a pressure value of approximately 1.0-1.5 N / cm²) to the meibomian gland area of ​​the patient's lower eyelid (usually the middle 1 / 3) for approximately 10-15 seconds, and record dynamic video of the secretion discharge using a high-definition camera (frame rate ≥30fps). The video should clearly capture the moment the secretion is squeezed out of the opening and its characteristics.

[0070] Step S2: Automatic scoring based on AI-driven four-dimensional features

[0071] The collected data is uploaded to the AI ​​engine via an encrypted network. The engine operates in parallel as follows:

[0072] 1. Data preprocessing and alignment:

[0073] All input images are uniformly scaled to 256×256 pixels.

[0074] Affine transformation registration was performed on the color and infrared images to make the eyelid margin line roughly horizontal and centered.

[0075] Keyframes were extracted from the dynamic video. Frames at 3, 6, 9, and 12 seconds after the press started were selected and combined with the baseline frame before the press to form a 5-frame time sequence.

[0076] Perform color normalization (for RGB images) and contrast-limited adaptive histogram equalization (for infrared images).

[0077] 2. Feature extraction and multi-task scoring:

[0078] The preprocessed data is input into a multi-task neural network optimized for eyelid anatomy. The core innovation of this network lies in its backbone network and region adaptive module.

[0079] Backbone Network: Employs an improved ResNet-50 architecture. In stages two through four, standard 3×3 convolutions are replaced with deformable strip convolutions (kernels of 1×9 and 9×1). The offset learning of the convolution kernels is guided by a pre-generated parabolic mask simulating the curvature of the eyelid, allowing the network's receptive field to adaptively extend along the meibomian gland orientation (approximately horizontal), greatly enhancing its ability to capture features of curved, elongated glandular structures.

[0080] Region Adaptive Module (MG-RAM):

[0081] The network first predicts a rough "glandular region probability map" on the feature map extracted from the backbone network.

[0082] Based on this probability map, dynamic gland contour-aware pooling is performed: The AI ​​comprehensive analysis engine first quickly generates a "gland region probability map," predicting which regions in the image are most likely to contain meibomian glands. Then, based on this probability map, a deformable sampling grid is generated. This grid becomes denser in high-probability regions (where glands are present) and sparser in low-probability regions (where there are no glands or background). Feature extraction (pooling) through this "intelligent" grid allows the model to focus more accurately and efficiently on effective information. By generating a deformable grid with denser grid points in regions with high gland probability and sparser points in missing regions, feature pooling is performed only on regions containing effective glands.

[0083] The pooled features will pass through a missing pattern classifier (which determines whether the glands in the region are normal, mild, moderate, or severe missing). The classification result will be used to dynamically adjust the feature response intensity of the decoders for subsequent tasks in the form of conditional modulation.

[0084] 3. Parallel calculation of four-dimensional scores:

[0085] (1) Mesothelial gland orifice obstruction grading criteria (MGO):

[0086] 0 points (normal) = no meibomian gland opening blockage;

[0087] 1 point (mild) = less than 1 / 3 of the meibomian gland opening is blocked;

[0088] 2 points (moderate) = 1 / 3 to 2 / 3 of the meibomian gland opening is blocked;

[0089] 3 points (severe) = more than 2 / 3 of the meibomian gland openings are blocked;

[0090] (2) Abnormal characteristics of meibomian gland secretions (MGS): Grading and scoring can be performed, with the highest score seen as the examination result, followed by analysis of lipid composition, proteomics and detection of inflammatory response factors in meibomian gland secretions.

[0091] Grading and scoring criteria:

[0092] 0 points (normal) = clear, transparent liquid discharge;

[0093] 1 point (mild) = cloudy liquid discharge;

[0094] 2 points (moderate) = cloudy, granular secretions;

[0095] 3 points (severe) = thick, toothpaste-like discharge;

[0096] (3) Abnormal facial gland drainage capacity (MGA): The number of facial gland openings that can drain facial secretions within the range of 5 facial gland openings on the outer side and center of the lower eyelash line.

[0097] Grading and scoring criteria:

[0098] 0 points (normal) = secretions are discharged from all 5 meibomian gland openings;

[0099] 1 point (mild) = secretions discharged from 3 or 4 meibomian gland openings;

[0100] 2 points (moderate) = secretions discharged from 1 or 2 meibomian gland openings;

[0101] 3 points (severe) = no meibomian gland openings to discharge secretions;

[0102] (4) Scoring criteria for meibomian gland absence area (MGL):

[0103] 0 points: No meibomian glands are missing, the glandular structure is intact, and the distribution is even;

[0104] 1 point: The proportion of meibomian gland absence is <1 / 3, that is, the absence area accounts for less than one-third of the total area of ​​the meibomian glands;

[0105] 2 points: The proportion of meibomian gland absence is 1 / 3 to 2 / 3, and the absence area accounts for one-third to two-thirds of the total area of ​​the meibomian gland;

[0106] 3 points: The proportion of meibomian gland absence is greater than 2 / 3, and the absence area exceeds two-thirds of the total area of ​​meibomian glands.

[0107] Step S3: Calculation and grading of total score

[0108] The four scores of MGO, MGS, MGA and MGL for the same eye are added together to obtain the comprehensive total score (MGS-Total, range 0-12 points).

[0109] 0-1 points: Considered as non-MGD or clinically insignificant changes.

[0110] 2-5 points: Diagnosed as mild MGD.

[0111] 6-9 points: Diagnosed as moderate MGD.

[0112] 10-12 points: Diagnosed as severe MGD.

[0113] Step S4: Treatment Follow-up and Prognostic Assessment

[0114] The AI-powered comprehensive analysis engine creates electronic records for each patient. When a patient returns for a follow-up visit, the engine automatically retrieves historical data for longitudinal comparison.

[0115] Macro-level assessment: Comparing the total score this time with the previous time, a decrease in the total score indicates an overall improvement in the condition.

[0116] Micro-analysis: Compare the changes in the scores of the four sub-items.

[0117] If MGO, MGS, and MGA improve but MGL remains unchanged, it suggests that the treatment (such as physical massage or hot compress) has effectively improved duct patency and secretion quality, but has not yet reversed glandular atrophy.

[0118] If the MGL score also improves (rare and requires long-term treatment), it suggests that the treatment may have promoted some degree of repair of the glandular structure.

[0119] This analysis can provide doctors with precise information to adjust treatment plans (such as strengthening anti-inflammatory treatment or considering intense pulsed light therapy).

[0120] The model training implementation method in this embodiment is as follows:

[0121] 1. Data Annotation: Collected multimodal image data from thousands of confirmed MGD patients and healthy volunteers. Annotation was independently completed by at least two senior ophthalmologists: delineated glandular regions pixel-wise on infrared images (morphological annotation); marked the status of each opening on color images; marked discharge events and secretion characteristics in videos; and provided expert scores across four dimensions for each data point.

[0122] 2. Training process:

[0123] Pre-training: The backbone network is pre-trained on a large natural image dataset (such as ImageNet).

[0124] Task-based training: The U-Net segmentation network is trained using morphologically labeled data, and the functional evaluation network is trained using functionally labeled data.

[0125] Joint fine-tuning: The segmentation network and the functional evaluation network are connected via MG-RAM modules to construct a complete multi-task network. End-to-end joint fine-tuning is performed using a weighted multi-task loss function. The loss function is:

[0126]

[0127] Lconsistency is used to constrain the logical consistency between "morphological loss rate" and "excretion capacity / secretory characteristics". For example, severely missing glands correspond to a significant decrease in excretion capacity, strengthening the pathological rationality of the model. LMGO measures the difference between the shape of the gland segmented by the AI ​​and the shape drawn by the doctor manually; LMGS measures the degree of consistency between the AI's judgment of the opening blockage and the doctor's judgment; LMGA represents the characteristics of the transverse secretions; and LMGL represents the accuracy of the judgment of excretion capacity. These represent the weighting factors. Because the four tasks have different difficulties and importance, setting these weighting factors tells the model which task to focus on optimizing during training. For example, if λ4 is set relatively large, the model will work harder to learn how to accurately segment gland morphology. These weight values ​​are optimal values ​​determined by AI engineers through extensive experimentation during model development.

[0128] 3. Optimization and Validation: Five-fold cross-validation is used to evaluate model performance. The Dice coefficient is used to evaluate segmentation accuracy, and accuracy and F1 score are used to evaluate the classification and scoring task. The final model's overall score on the independent test set should achieve a Kappa value > 0.85 in agreement with expert scores.

[0129] The AI ​​comprehensive analysis engine output example is as follows:

[0130] After the process is complete, the doctor's terminal will display the following report:

[0131] Patient information: ID, name, examination date.

[0132] Monocular scoring results:

[0133] MGO: 2 points (moderate opening blockage)

[0134] MGS: 3 points (Discharge is toothpaste-like)

[0135] MGA: 2 points (moderately decreased excretion capacity)

[0136] MGL: 2 points (moderate glandular absence, absence rate 35%)

[0137] Overall score: 9 points (Moderate MGD)

[0138] Visualization results:

[0139] The outline of the glands segmented by AI is overlaid on the infrared image (green represents normal glands, and red represents missing areas).

[0140] The color image highlights the gland openings that are identified as blocked.

[0141] The moment of successful discharge is marked in the dynamic video.

[0142] Historical comparison chart: The trend of changes in the four scores and total score of this inspection compared with previous inspections is shown in the form of a line graph.

[0143] Treatment recommendations: Based on the sub-scores, the AI ​​comprehensive analysis engine may suggest that "the main problems in this case are abnormal secretion characteristics and obstruction of the opening. It is recommended to strengthen meibomian gland massage and local anti-inflammatory treatment."

[0144] During the offset learning process, it is constrained by the eyelid curvature prior map, and a curved mask of the same size as the input image is generated in advance (the upper eyelid is an upwardly convex parabola, and the lower eyelid is downwardly concave, such as...). Figure 2 As shown in the diagram, this mask is then injected as a positional code into a deformable convolutional offset prediction network, causing the sampling points to be preferentially stretched along the long axis of the eyelid (approximately horizontal), forming a "serpentine receptive field following the eyelid margin." Compared to traditional 3×3 square convolutions, with the same computational cost, the ability to capture the integrity of curved glands is improved by 370% (internal ablation experiment), such as... Figure 3 As shown.

[0145] Then, the Tri-Layer Anatomy AttentionGate (TAAG) adds a lightweight branch after each layer's output in the backbone, predicting three soft masks: the eyelid margin area, the conjunctival area, and the skin area (corresponding to the anterior, middle, and posterior tarsal plates anatomically). The process is as follows: Figure 4 As shown, these three masks are used to perform weighted modulation on the feature map, which automatically enhances high-frequency edge information near the eyelid margin, enhances texture contrast in the conjunctival area, and suppresses irrelevant noise in the skin area.

[0146] Traditional ViT / Swin uses standard sinusoidal positional encoding. This technical solution additionally introduces "upper and lower eyelid label images" (all 1s for the upper eyelid, all 0s for the lower eyelid, and 0.5s for the rest), converting them into learnable embeddings and adding them to the original positional encoding. This allows the network to know from the first token whether it belongs to the upper or lower eyelid, avoiding confusion when flipping eyelid images. Traditional multi-task networks (such as Mask R-CNN + classification head, UBER-Net, MTI-Net) typically directly connect shared features to each task head, ignoring the fact that meibomian gland images have "extremely irregular morphology of regions of interest, high gland loss rate, and complex background noise," leading to severe interference between different tasks.

[0147] In this technical solution, the probability map is used as a "soft attention mask" to guide subsequent RoI feature pooling. This allows the pooling grid points to automatically shrink towards the remaining glands in areas with missing glands and automatically stretch in areas with excessive gland distortion, forming an "adaptive grid that deforms with the true contour of the glands." The difference from Deformable RoI Pooling is that the latter only learns geometric deformation in its offset calculation, while the offset in this solution is explicitly constrained by the "probability of gland presence," avoiding pooling into large blank areas. Furthermore, before the shared features enter each dedicated decoder, a 4-class auxiliary head (normal, mildly missing, moderately missing, severely missing / completely occluded) is added. This converts the one-hot vector of the missing pattern into an embedding and performs conditional feature modulation with the region features (similar to a FiLM layer). This allows the subsequent segmentation head to automatically "lower the segmentation threshold and amplify the residual gland signal" in cases of severe missing glands, while the classification head "suppresses overconfidence." This is equivalent to letting the network first determine "whether there are still glands in this region" before deciding how to segment and score, completely solving the problem of missegmentation when glands are almost completely absent in patients with severe dry eye.

[0148] Traditional multi-task methods use uncertainty weighting or GradNorm. This invention proposes an "adaptive task weighting based on gland area proportion": It calculates the percentage of gland pixels (A) in the current image in real time. If A < 5%, it automatically reduces the segmentation loss weight to 0.1 and increases the image-level scoring loss weight to 0.9, preventing the segmentation task from dragging down overall convergence. The aforementioned MG-RAM module is placed after the shared backbone and before each dedicated decoder, completely different from any existing multi-task hard or soft parameter sharing scheme. It is specifically designed for scenarios with "extreme meibomian gland absence and highly variable morphology."

[0149] In this embodiment, a simplified AI engine can be provided for primary healthcare institutions with limited computing resources, while the functional scores (MGO, MGS, MGA) are performed using a lightweight MobileNet network.

[0150] The trained AI model can also be packaged into software and installed on a local graphics workstation, running without a network and ensuring data privacy.

[0151] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-dimensional quantitative assessment method combining meibomian gland morphology and function, characterized in that, Includes the following steps: Multimodal information acquisition steps: Acquire patient eyelid multimodal data including color static images of the eyelid margin, infrared images of the meibomian glands, and dynamic video of secretions being discharged upon pressure; The four-dimensional automatic scoring process involves inputting the multimodal data into an AI analysis model, which then outputs quantitative scores across four dimensions in parallel or sequentially, including: a. Meibomian gland orifice abnormality score; b. Score of meibomian gland secretion characteristics; c. Meibomian gland excretion capacity score; d. Meibomian gland absence area score, which is obtained by calculating the absence rate and mapping it based on the gland segmentation results in the infrared image; Comprehensive assessment steps: The scores of the four dimensions are added together to obtain the total score of meibomian gland function and morphology, and the meibomian gland dysfunction is graded according to the total score range; Treatment efficacy tracking steps: Compare the scores of the four dimensions and the overall score of the same patient at different time points to assess the treatment effect and analyze the improvement dimensions.

2. The multi-dimensional quantitative assessment method combining meibomian gland morphology and function according to claim 1, characterized in that, The AI ​​analysis model includes: The anatomically guided backbone network has at least some of its convolutional layers using deformable strip convolutions, and the convolution offset is constrained by the eyelid curvature prior map. The meibomian gland region adaptive module is used for dynamic contour-aware pooling based on the gland presence probability map and conditional modulation of features based on missing modes. The four-dimensional scoring subnetwork is used to perform opening abnormality identification, secretion morphology classification, expulsion event detection, and gland segmentation and missing rate calculation.

3. The multi-dimensional quantitative assessment method combining meibomian gland morphology and function according to claim 2, characterized in that, The meibomian gland region adaptive module includes: Glandular contour probability map prediction unit; Pooling units that generate an adaptive sampling grid based on the probability graph; A missing pattern classifier is used to output the classification of glandular absence degree and to conditionally modulate features for subsequent tasks.

4. The multi-dimensional quantitative assessment method combining meibomian gland morphology and function according to claim 1, characterized in that, The scoring steps for the area of ​​meibomian gland absence include: Use U-Net or a neural network with an improved U-Net architecture to perform gland pixel-level segmentation in infrared images; The meibomian gland absence rate is calculated using the formula: (Total palpebral conjunctival area / Total gland area) × 100%. Based on the preset missing rate threshold range, the missing rate is mapped to a score of 0-3.

5. The multi-dimensional quantitative assessment method combining meibomian gland morphology and function according to claim 1, characterized in that, The secretion phenotype scores were obtained based on dynamic video analysis, specifically including: Temporal convolutional networks or ConvLSTM are used to extract spatiotemporal features of color, transparency and morphological changes during the secretion process; The extracted features are classified as clear, cloudy, granular, or toothpaste-like, and mapped to a score of 0-3.

6. The multi-dimensional quantitative assessment method combining meibomian gland morphology and function according to claim 1, characterized in that, The discharge capacity score is obtained based on dynamic video analysis, specifically including: The gland openings that successfully expel secretions during the pressing process are identified using optical flow or trajectory tracking technology. The proportion of successfully discharged openings to the total number of pressurized openings was statistically analyzed and mapped to a score of 0-3.

7. The multi-dimensional quantitative assessment method combining meibomian gland morphology and function according to claim 1, characterized in that, The method further includes: During the training phase of the AI ​​analysis model, a multi-task loss function is used for joint optimization. The loss function includes loss terms for each dimension of scoring tasks and a logical consistency loss term, which is used to constrain the pathological rationality between morphological deficiencies and functional scores.

8. A multi-dimensional quantitative assessment system combining meibomian gland morphology and function for implementing the method of any one of claims 1-7, characterized in that, include: A multimodal acquisition device is used to acquire color static images of the eyelid margin, infrared images of the meibomian glands, and dynamic videos of secretion discharge. The AI ​​analysis engine is deployed with the AI ​​analysis model as described in claims 2-3, which is used to receive the multimodal data and output a four-dimensional score and a comprehensive total score; The report generation unit integrates scoring results, visualizations, and historical comparison information to generate a structured evaluation report.

9. The multi-dimensional quantitative assessment system combining meibomian gland morphology and function according to claim 8, characterized in that, The AI ​​analysis engine further includes: The data preprocessing module is used for spatial registration, size normalization, image enhancement, and video keyframe extraction of multimodal images; The model interpretation and visualization module is used to generate gland segmentation overlay maps, opening anomaly annotation maps, and high-risk area heat maps.

Images