Deep learning based meibomian gland image segmentation and morphological quantification method and device

Abstract

The application relates to a deep learning-based meibomian gland image segmentation and morphological quantification method and device, which comprises the following steps: acquiring one of an upper eyelid image or a lower eyelid image collected by an infrared imaging device; inputting the eyelid image into a pre-trained multi-task learning network to output a segmentation result; the segmentation result comprises an instance segmentation mask of each independent meibomian gland and a region segmentation mask of a meibomian gland; based on the segmentation result, multi-dimensional morphological parameters of each independent meibomian gland are extracted, and a meibomian gland classification label of the eyelid image is determined; the multi-dimensional morphological parameters are used to represent the morphological structure, distribution characteristics and functional state of the meibomian gland; the meibomian gland classification label is an upper meibomian gland or a lower meibomian gland; according to the meibomian gland classification label and the multi-dimensional morphological parameters, a corresponding meibomian gland state evaluation result is generated. Comprehensive coverage of upper and lower eyelid meibomian glands, individualized instance segmentation, and automatic extraction and classification evaluation of multi-dimensional morphological parameters are realized.

Description

Technical Field

[0001] This application belongs to the field of image processing and computer vision technology, and in particular relates to a method and apparatus for meibomian gland image segmentation and morphological quantization based on deep learning. Background Technology

[0002] The meibomian glands (MG) within the tarsal plate of the eyelid are composed of vertical sebaceous glands. Their function is to secrete oil to reduce tear evaporation and protect the ocular surface. Meibomian gland dysfunction (MGD) is a common, multifactorial, chronic, diffuse MG disease that can cause eye discomfort, tear problems, inflammation, and decreased vision, and is a major contributing factor to dry eye. MGD is prevalent worldwide, with a particularly high prevalence in Asian populations.

[0003] Clinically, the diagnosis of MGD largely relies on the analysis of the appearance of the eyelids and glands. Non-invasive infrared tarsal plate imaging is a routine technique for assessing MG morphology; however, this method is heavily dependent on expert experience, highly subjective, difficult to quantify early glandular damage, and exhibits significant inter-observer variability, affecting the accurate diagnosis of MGD. The large discrepancies in the prevalence of MGD in the literature are also related to the lack of standardized diagnostic criteria. Therefore, developing a non-invasive, rapid, accurate, objective, and quantifiable diagnostic system for MGD is of significant clinical importance.

[0004] In recent years, artificial intelligence (AI), especially deep learning (DL) technology, has provided new avenues for automated meibomian gland (MG) analysis. International research has utilized AI technology to train on expert-annotated datasets to automatically identify and segment meibomian glands and calculate the MG loss rate. However, existing automated meibomian gland image analysis methods still face the following technical challenges: (1) Most existing automated methods only analyze images of the upper eyelid and fail to fully cover the lower eyelid, resulting in incomplete assessment results and an inability to fully reflect the overall condition of the patient's meibomian glands. At the same time, existing methods can usually only output the overall outline or regional segmentation results of the meibomian glands. They are based on a rough estimate of the missing area of ​​the glands in the meibomian glands and cannot achieve instance-level segmentation of each individual meibomian gland. It is difficult to perform fine analysis and quantitative assessment of individual glands and it is easy to underestimate early atrophy.

[0005] (2) Existing automated methods extract relatively simple morphological parameters, mainly focusing on macroscopic indicators such as gland area or outline loss rate. They lack comprehensive quantification of multidimensional morphological parameters such as gland length, width distribution, degree of tortuosity, and functional status, making it difficult to capture subtle pathological changes of MGD early and sensitively.

[0006] (3) Existing automated methods have high requirements for the quality of input images and their processing procedures are complex. They usually require tedious, separate processing of interference factors such as reflections and bright artifacts in the image beforehand in order to obtain a relatively ideal segmentation effect, which increases the complexity of the operation. In addition, since the meibomian gland infrared image acquisition equipment used abroad (such as the LipiView II ocular surface interferometer) is different from the mainstream image acquisition equipment used in China (such as the Keratograph 5M), existing models trained on foreign equipment data have poor recognition and segmentation effects when applied to images acquired by domestic equipment, making it difficult to directly meet the actual clinical needs.

[0007] To address the technical problems of traditional automated meibomian gland image analysis methods, such as incomplete evaluation scope, lack of individual gland instance segmentation capability, limited morphological parameters, and cumbersome image artifact processing procedures, no effective solution has yet been proposed. Summary of the Invention

[0008] In view of the shortcomings of the prior art, the purpose of the invention is to provide a method and device for meibomian gland image segmentation and morphological quantification based on deep learning. By constructing a multi-task learning network that can process upper and lower eyelid images simultaneously, instance-level segmentation of each independent meibomian gland is achieved. Based on this, multi-dimensional morphological parameters are extracted, and the upper and lower eyelids are automatically identified and classified, ultimately generating a comprehensive and refined state assessment result.

[0009] In a first aspect, this application proposes a deep learning-based method for meibomian gland image segmentation and morphological quantification, comprising: acquiring an eyelid image captured by an infrared imaging device, wherein the eyelid image is either an upper eyelid image or a lower eyelid image; inputting the eyelid image into a pre-trained multi-task learning network and outputting a segmentation result; the segmentation result includes an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; based on the segmentation result, extracting multidimensional morphological parameters for each independent meibomian gland and determining a meibomian gland classification label for the eyelid image; the multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional state of the meibomian gland; the meibomian gland classification label is either an upper meibomian gland or a lower meibomian gland; and generating a corresponding meibomian gland state evaluation result based on the meibomian gland classification label and the multidimensional morphological parameters.

[0010] According to a second aspect of the present disclosure, a storage medium is provided, the storage medium including a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0011] According to a third aspect of the present disclosure, a deep learning-based meibomian gland image segmentation and morphological quantization device is provided, comprising: an image acquisition module for acquiring eyelid images collected by an infrared imaging device, wherein the eyelid images are either upper eyelid images or lower eyelid images; an image segmentation module for inputting the eyelid images into a pre-trained multi-task learning network and outputting segmentation results; the segmentation results include an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; a parameter extraction module for extracting multidimensional morphological parameters for each independent meibomian gland based on the segmentation results and determining a meibomian gland classification label for the eyelid image; the multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional state of the meibomian gland; the meibomian gland classification label is either upper meibomian gland or lower meibomian gland; and a state evaluation module for generating corresponding meibomian gland state evaluation results based on the meibomian gland classification label and the multidimensional morphological parameters.

[0012] According to a fourth aspect of the present disclosure, a deep learning-based meibomian gland image segmentation and morphological quantization apparatus is provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions to perform the following processing steps: acquiring an eyelid image collected by an infrared imaging device, wherein the eyelid image is either an upper eyelid image or a lower eyelid image; inputting the eyelid image into a pre-trained multi-task learning network and outputting a segmentation result; the segmentation result including an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; based on the segmentation result, extracting multidimensional morphological parameters for each independent meibomian gland and determining a meibomian gland classification label for the eyelid image; the multidimensional morphological parameters being used to characterize the morphological structure, distribution characteristics, and functional state of the meibomian gland; the meibomian gland classification label being either an upper meibomian gland or a lower meibomian gland; and generating a corresponding meibomian gland state evaluation result based on the meibomian gland classification label and the multidimensional morphological parameters.

[0013] The beneficial effects of this application are as follows: (1) This application obtains images of the upper or lower eyelid and uses a multi-task learning network to output the instance segmentation mask of each independent meibomian gland and the region segmentation mask of the tarsal plate, thereby achieving comprehensive coverage and individualized precise segmentation of the meibomian glands of the upper and lower eyelids. This overcomes the shortcomings of existing technologies that focus on the upper eyelid and cannot achieve instance-level segmentation, and provides a basis for comprehensively assessing the condition of the patient's meibomian glands.

[0014] (2) Based on the segmentation results, this application extracts multidimensional morphological parameters of each independent meibomian gland, including but not limited to gland length, average width, 10th and 90th percentile width, diameter deformation index, tortuosity, individual meibomian gland parenchyma ratio, meibomian gland atrophy rate, meibomian gland density and signal intensity index, realizing comprehensive quantification of meibomian gland morphology, distribution characteristics and functional status, and enabling earlier and more sensitive capture of subtle pathological changes in MGD.

[0015] (3) This application integrates the processing of reflective highlight artifacts into the model by setting a highlight area detection branch and an image inpainting branch in the multi-task learning network, realizing "one-step" non-destructive repair. It does not require cumbersome preprocessing outside the model, simplifies the operation process, and improves diagnostic efficiency.

[0016] (4) This application automatically determines the meibomian gland classification label (upper meibomian gland or lower meibomian gland) of the eyelid image based on the segmentation results, and generates targeted status assessment results by combining multidimensional morphological parameters, thereby realizing the differential analysis of the upper and lower eyelids and improving the accuracy and clinical reference value of the assessment results.

[0017] Therefore, this application achieves comprehensive coverage of the meibomian glands of the upper and lower eyelids, individualized instance segmentation, and automated extraction and classification evaluation of multidimensional morphological parameters through an end-to-end deep learning model. This solves the technical problems of existing automated methods, such as incomplete evaluation scope, lack of individual gland instance segmentation capability, limited morphological parameters, and cumbersome external processing of image artifacts. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Throughout the drawings, the same reference numerals denote the same components. Obviously, the drawings described below are merely some embodiments described in this application, and those skilled in the art can obtain other drawings based on these drawings.

[0019] Figure 1 This is a hardware structure block diagram of a computing device for implementing the method described in Embodiment 1 of this disclosure; Figure 2 This is a flowchart of the deep learning-based meibomian gland image segmentation and morphological quantization method according to Embodiment 1 of this application; Figure 3 This is an overall framework diagram of the meibomian gland image segmentation and morphological quantization method based on deep learning as described in Embodiment 1 of this application; Figure 4This is a schematic diagram of infrared meibomian gland image annotation according to Embodiment 1 of this application, in which the annotation method of the meibomian region, a single meibomian gland, a highlighted area and the smallest overall area of ​​the meibomian gland is shown from left to right; Figure 5 This is a schematic diagram of the segmentation result of an upper eyelid image acquired by the infrared imaging device described in Embodiment 1 of this application, which carries an instance segmentation mask for each independent meibomian gland. Figure 6 This is a schematic diagram of the segmentation result of a region segmentation mask carrying the tarsal plate in an upper eyelid image acquired by the infrared imaging device described in Embodiment 1 of this application. Figure 7 This is a schematic diagram of the segmentation result of an image of the lower eyelid acquired by the infrared imaging device described in Embodiment 1 of this application, which carries an instance segmentation mask for each independent meibomian gland. Figure 8 This is a schematic diagram of the segmentation result of a region segmentation mask carrying the tarsal plate in an image of the lower eyelid acquired by the infrared imaging device described in Embodiment 1 of this application. Figure 9 This is a schematic diagram of the meibomian gland image segmentation and morphological quantization device based on deep learning according to Embodiment 2 of this application; Figure 10 This is a schematic diagram of the meibomian gland image segmentation and morphological quantization device based on deep learning as described in Embodiment 3 of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The core parameters involved in this application are defined as follows: 1. Meibomian gland (MG): meibomian gland; 2. Tarsal plate: the tarsal plate of the eyelid; 3. Tarsal plate area (PA): Area of ​​the tarsal plate; 4. Contoured meibomian gland area (CGA): Area of ​​the meibomian glands in the circumference of ... 5. Individual meibomian gland area (IGA): The area of ​​the meibomian glands in an individual. 6. Contour meibomian gland area ratio = (CGA / PA) × 100%; 7. Individual meibomian gland area ratio = (IGA / PA) × 100%; 8. Glandular imaging index = (Average gray value of IG / Average gray value of non-glandular area) × 100% Example 1

[0023] According to this embodiment, a method embodiment of meibomian gland image segmentation and morphological quantization based on deep learning is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than that shown here.

[0024] The method embodiments provided in this example can be executed on a server or similar computing device. Figure 1 A hardware block diagram of a computing device for implementing a deep learning-based method for meibomian gland image segmentation and morphological quantization is shown. Figure 1 As shown, a computing device may include one or more processors (processors may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), a memory for storing data, a transmission device for communication functions, and an input / output interface. The memory, transmission device, and input / output interface are connected to the processor via a bus. In addition, it may also include a display, keyboard, and cursor control device connected to the input / output interface. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a computing device may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0025] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element in a computing device. As involved in the embodiments of this disclosure, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0026] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the deep learning-based meibomian gland image segmentation and morphological quantization method in this embodiment of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the deep learning-based meibomian gland image segmentation and morphological quantization method of the aforementioned application. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the computing device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0027] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the computing device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0028] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of the computing device.

[0029] It should be noted here that, in some optional embodiments, the above... Figure 1 The computing device shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computing devices.

[0030] Under the above operating environment, according to the first aspect of this embodiment, a deep learning-based method for meibomian gland image segmentation and morphological quantization is provided. Figure 2 A flowchart illustrating the method is shown. Figure 3 This diagram illustrates the overall framework of the method. The complete process from input image to final meibomian gland status assessment results can be understood by referring to this diagram. (Reference) Figure 2 and Figure 3 As shown, the method includes: S202: Acquire an eyelid image captured by an infrared imaging device, wherein the eyelid image is either an upper eyelid image or a lower eyelid image; S204: Input the eyelid image into a pre-trained multi-task learning network and output the segmentation result; the segmentation result includes an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; S206: Based on the segmentation results, extract the multidimensional morphological parameters of each independent meibomian gland and determine the meibomian gland classification label of the eyelid image; the multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics and functional status of the meibomian gland; the meibomian gland classification label is upper meibomian gland or lower meibomian gland; S208: Generate the corresponding meibomian gland status assessment result based on the meibomian gland classification label and the multidimensional morphological parameters.

[0031] In this embodiment, during clinical application, when an objective, comprehensive, and quantitative assessment of meibomian gland dysfunction is required, eyelid images acquired by an infrared imaging device are first obtained. These eyelid images can be either upper or lower eyelid images (corresponding to step S202). Specifically, in clinical applications, Keratograph 5M, LipiView II, or other ophthalmic devices with infrared illumination imaging capabilities can be used to acquire infrared images of the patient's eyelids. The acquired images can be either upper or lower eyelid images. For example, in rapid diagnostic scenarios, the acquired eyelid images can be directly input into the system; in large-scale physical examinations, a large number of upper and lower eyelid images can be acquired and processed in batches; in longitudinal monitoring of treatment efficacy, images can be acquired at multiple time points before, during, and after treatment to track parameter changes. This step supports the separate acquisition of both types of images, laying the foundation for a comprehensive assessment of the patient's meibomian gland status.

[0032] Then, the eyelid image is input into a pre-trained multi-task learning network, which outputs segmentation results. These results include an instance segmentation mask for each individual meibomian gland and a region segmentation mask for the tarsal plate (corresponding to step S204). Specifically, in this step, the pre-trained multi-task learning network can perform pixel-level segmentation analysis on the input raw infrared eyelid image and output two types of segmentation results: first, an instance segmentation mask for each individual meibomian gland, which distinguishes and labels each meibomian gland in the image as an independent entity, rather than simply identifying the overall region of all glands; and second, a region segmentation mask for the tarsal plate, which accurately delineates the complete region of the tarsal bone (tarsal plate) in the eyelid.

[0033] During the model training phase, the above multi-task learning network is constructed and trained through the following steps: First, a standardized dataset is constructed, containing both upper and lower eyelid images. Then, image preprocessing and data augmentation are performed on the dataset. Image preprocessing includes contrast enhancement, noise suppression, reflection area correction, and image normalization. Data augmentation includes color perturbation, random illumination adjustment, horizontal flipping, contrast balancing, and scale perturbation to improve the model's generalization ability to images from different devices, under different exposure conditions, and with varying eyelid flipping quality.

[0034] Preferably, the contrast enhancement employs the Contrast Limited Adaptive Histogram Equalization (CLAHE) method. CLAHE divides the image into multiple local regions, performs local histogram equalization on each, and limits local noise amplification by setting a clipping threshold, thereby improving the grayscale contrast between the meibomian glands and the background tissue. In infrared images, the grayscale difference between the glands and the background (tarsal tissue) is often small, and due to limitations in hardware imaging principles, the image may contain significant random noise. Contrast Limited Adaptive Histogram Equalization (CLAHE) performs exceptionally well in this type of image enhancement. CLAHE divides the image into non-overlapping small regions (Tiles), performs local histogram equalization, and introduces a clipping threshold to prevent excessive noise amplification. Experimental data shows that CLAHE processing can significantly improve the accuracy of medical image classification and segmentation, increasing the recognition rate by 3% to 4% on certain ophthalmic datasets.

[0035] Preferably, the noise suppression in step S2 employs a bilateral filter or a Butterworth low-pass filter. The bilateral filter simultaneously considers pixel spatial distance and grayscale similarity, reducing random noise in the image while preserving meibomian gland edge information. Specifically, to further smooth the image and preserve gland edge information, the technical solution typically introduces a bilateral filter or a Butterworth low-pass filter. While smoothing noise, the bilateral filter effectively protects the boundary between the meibomian gland and surrounding tissue from blurring by considering pixel spatial proximity and intensity similarity.

[0036] Preferably, the reflective area correction adopts a high reflectivity mask replacement algorithm based on threshold detection; when there are teardrop reflections or liquid spot highlights in the image, the reflective area is extracted by threshold segmentation, and the reflective area is replaced by the mean or interpolation result of the surrounding neighboring pixels to reduce false positives during the segmentation process.

[0037] Then, the acquired raw infrared images of meibomian glands were manually annotated. The annotations included the meibomian region, meibomian glands, highlighted areas, and the smallest overall area of ​​the meibomian glands, to form training, validation, and test datasets. Figure 4 As shown, from left to right, the annotation methods for the meibomian region, a single meibomian gland, a highlighted area, and the minimum overall range of the meibomian gland are illustrated. Because the highlighted area is included as an independent annotation category in the training data during the model training phase, the model can autonomously learn to identify reflective highlight artifacts in the image and the true structural features of the surrounding glandular tissue. Therefore, in the application phase, no external artifact processing is required on the acquired eyelid images for direct analysis. It should be noted that the minimum overall range of the meibomian gland refers to the entire meibomian gland envelope region, which is the area obtained by calculating the minimum circumscribed convex polygon or minimum circumscribed rectangle of all meibomian glands in the same eyelid image. This region is used to assist in calculating the meibomian gland atrophy rate based on the contour envelope method, comparing it with the atrophy rate based on individual meibomian gland instance segmentation to evaluate the sensitivity advantage of individualized segmentation methods in early lesion detection.

[0038] Preferably, the meibomian plate region and meibomian gland region are annotated at the full pixel level to improve the segmentation model’s ability to learn the contours and boundary details of the meibomian glands. In some training scenarios, sparse annotation or graffiti annotation can also be used as auxiliary supervision information to reduce annotation costs.

[0039] Next, the multi-task learning network was trained with a large amount of labeled data covering both the upper and lower eyelids, enabling it to learn the ability to accurately identify and segment the tarsal plate region and each individual meibomian gland from complex backgrounds. Since the training data covers both the upper and lower eyelids, the network can achieve stable and accurate segmentation for both types of images when applied.

[0040] Therefore, in the application phase, the upper eyelid image or lower eyelid image obtained in step S202 is input into the trained multi-task learning network, and the multi-task learning network can automatically output the above segmentation results. Figure 5 and Figure 6 The diagrams show the segmentation results of the upper eyelid image with instance segmentation masks carrying each individual meibomian gland and the segmentation results of the region segmentation mask carrying the meibomian gland. Figure 7 and Figure 8 The diagrams show the segmentation results of lower eyelid images with instance segmentation masks carrying each individual meibomian gland and the segmentation results with region segmentation masks carrying the tarsal plates, respectively. Through the aforementioned multi-task learning network, comprehensive coverage and individualized precise segmentation of the meibomian glands in both the upper and lower eyelids are achieved. This captures the actual pixel proportion of each individual gland, providing a precise data foundation for subsequent refined quantitative analysis.

[0041] After completing step S204, based on the segmentation results, multidimensional morphological parameters of each independent meibomian gland are extracted, and the meibomian gland classification label of the eyelid image is determined (corresponding to step S206). Specifically, this step automatically calculates and extracts a series of quantitative parameters that characterize the morphological structure, distribution characteristics, and functional status of the meibomian glands, based on the instance segmentation mask of each independent meibomian gland and the region segmentation mask of the tarsal plate output in step S204. These multidimensional morphological parameters include, but are not limited to: gland length, average width, 10th percentile width, 90th percentile width, diameter deformation index, tortuosity, individual meibomian gland parenchyma percentage, meibomian gland atrophy rate, meibomian gland density, and signal intensity index. These parameters can quantitatively assess the health status of the meibomian glands from different dimensions (the specific calculation methods and clinical significance of each parameter will be described in detail below), providing a refined quantitative basis for subsequent differential status assessment. At the same time, this step also automatically determines whether the current eyelid image belongs to the upper or lower eyelid based on the region segmentation mask of the tarsal plate in the segmentation results, and generates the corresponding meibomian gland classification label. The determination of this classification label provides a basis for subsequent differentiated status assessments for different eyelid locations.

[0042] Finally, based on the meibomian gland classification labels and the multidimensional morphological parameters, a corresponding meibomian gland status assessment result is generated (corresponding to step S208). Specifically, this step integrates the multidimensional morphological parameters extracted in step S206 with the determined upper and lower eyelid classification labels to generate a structured meibomian gland status assessment result. For example, a standardized Meiboscore score (0-3 or 0-6) can be automatically generated based on the meibomian gland atrophy rate, and combined with other parameters (such as the 10th percentile width reflecting the degree of atrophy, the 90th percentile width reflecting the degree of expansion, the tortuosity reflecting compensatory growth, and the signal intensity index reflecting the functional status) to provide a more detailed morphological description. Simultaneously, based on the upper and lower eyelid classification labels, the meibomian gland status of the upper and lower eyelids can be presented separately in the assessment report, enabling differentiated analysis. In clinical applications, this assessment result can serve as an important reference for doctors to diagnose MGD, assess the severity of the condition, and formulate treatment plans. Furthermore, in longitudinal monitoring scenarios, by performing the above analysis on images collected from the same patient at different time points, the changing trends of various morphological parameters can be output. For example, an increase in the 10th percentile width after treatment suggests that the narrow area of ​​the gland may be repaired, while a decrease in the 90th percentile width suggests that the abnormal expansion may be improved. This provides objective quantitative data for evaluating the efficacy of interventions such as hot compresses, intense pulsed light therapy, and meibomian gland massage.

[0043] As described in the background section, most existing automated meibomian gland (MGD) image analysis methods only analyze upper eyelid images, failing to comprehensively cover the lower eyelid. This results in incomplete assessments and an inability to fully reflect the overall condition of the patient's meibomian glands. Furthermore, existing methods typically only output the overall outline or region segmentation of the meibomian glands, failing to achieve instance-level segmentation of each individual gland, making it difficult to perform refined analysis and quantitative assessment of individual glands. Additionally, the morphological parameters extracted by existing automated methods are relatively limited, mainly focusing on macroscopic indicators such as gland area or outline loss rate, lacking comprehensive quantification of multidimensional morphological parameters such as gland length, width distribution, degree of tortuosity, and functional status. This makes it difficult to capture subtle pathological changes in MGD early and sensitively.

[0044] In view of this, this application first acquires images of the upper or lower eyelid and then uses a multi-task learning network to output instance segmentation masks for each independent meibomian gland and region segmentation masks for the tarsal plate. This achieves comprehensive coverage and individualized, precise segmentation of the meibomian glands in both the upper and lower eyelids, overcoming the shortcomings of existing technologies that primarily focus on the upper eyelid and cannot achieve instance-level segmentation. This provides a foundation for a comprehensive assessment of the patient's meibomian gland status. Then, based on the segmentation results, multidimensional morphological parameters are extracted for each independent meibomian gland. This enables a comprehensive quantification of the meibomian gland's morphological structure, distribution characteristics, and functional status, allowing for earlier and more sensitive detection of subtle pathological changes in MGD. Finally, based on the meibomian gland classification labels and multidimensional morphological parameters, corresponding meibomian gland status assessment results are generated, enabling differentiated analysis of the upper and lower eyelids and improving the accuracy and clinical reference value of the assessment results. This effectively solves the technical problems of traditional automated meibomian gland image analysis methods, such as incomplete assessment scope, lack of individual gland instance segmentation capabilities, and limited morphological parameters.

[0045] Optionally, the multi-task learning network includes a unified segmentation model and an adaptive interactive information fusion module; and the eyelid image is input into the pre-trained multi-task learning network to output segmentation results, including: inputting the eyelid image into the unified segmentation model to output a region segmentation mask for the meibomian glands, a region segmentation mask for the tarsal plates, and a boundary segmentation mask between adjacent meibomian glands; inputting the region segmentation mask for the meibomian glands and the boundary segmentation mask between adjacent meibomian glands into the adaptive interactive information fusion module for feature fusion, and outputting an instance segmentation mask for each independent meibomian gland.

[0046] Specifically, in combination Figure 3 As shown, the multi-task learning network used in this application is structurally divided into two core parts: a unified segmentation model and an adaptive interactive information fusion module. The unified segmentation model is responsible for performing preliminary multi-task segmentation analysis on the input eyelid image. This model does not output only a single segmentation result, but simultaneously outputs three different types of segmentation masks: first, a meibomian gland region segmentation mask, used to mark the overall area occupied by all meibomian glands in the image; second, a tarsal plate region segmentation mask, used to mark the complete area of ​​the tarsal bone (tarsal plate) in the eyelid; and third, a boundary segmentation mask between adjacent meibomian glands, used to identify and mark the narrow gap boundaries between adjacent glands. By simultaneously outputting these three masks, the unified segmentation model provides rich basic information for subsequent refined processing. It should be noted that during the model training phase, the unified segmentation model is trained using a large amount of data labeled with tarsal plate regions, meibomian glands, and the boundaries of adjacent glands, enabling it to accurately identify the boundaries between gland regions.

[0047] Then, the meibomian gland region segmentation mask output by the unified segmentation model and the boundary segmentation mask between adjacent meibomian glands are input into the adaptive interactive information fusion module. This module, based on a reverse attention mechanism, dynamically adjusts the fusion ratio of region features and boundary features, embedding boundary information as a constraint into the region segmentation result. This effectively separates adjacent glands that were originally stuck together and difficult to distinguish at the pixel level, ultimately outputting a segmentation mask for each independent meibomian gland instance. Through this structural design, not only can the overall region of the meibomian gland be identified, but each independent gland instance can also be accurately distinguished, providing a foundation for subsequent extraction of morphological parameters such as length, width, and curvature for each gland individually.

[0048] Therefore, by setting a unified segmentation model to simultaneously output region segmentation masks and boundary segmentation masks, and using an adaptive interactive information fusion module for feature fusion, effective separation of adhesiomyocytes is achieved. This overcomes the shortcomings of existing technologies that cannot achieve instance-level segmentation of each independent meibomian gland, and provides a data foundation for individualized and refined quantitative analysis.

[0049] Optionally, the unified segmentation model includes an encoder, a striped hybrid attention module, an attention gating module, a decoder, a boundary detection branch, and multiple segmentation heads. Furthermore, inputting the eyelid image into the unified segmentation model and outputting a region segmentation mask for the meibomian glands, a region segmentation mask for the tarsal plate, and a boundary segmentation mask between adjacent meibomian glands includes: inputting the eyelid image into the encoder for feature extraction, and inputting the features output by the encoder into the striped hybrid attention module to extract long-distance dependent features using horizontal and vertical long strip convolution kernels; inputting the features output by the striped hybrid attention module into the attention gating module and the boundary detection branch, respectively; identifying the narrow gap boundary between adjacent meibomian glands through the boundary detection branch and outputting a boundary segmentation mask between adjacent meibomian glands; inputting the features output by the attention gating module into the decoder for upsampling and feature reconstruction, and inputting the features output by the decoder into the multiple segmentation heads to output region segmentation masks for the meibomian glands and tarsal plate, respectively.

[0050] Specifically, in combination Figure 3 As shown, the unified segmentation model in this application employs a multi-task learning strategy in its structure, aiming to simultaneously solve three sub-tasks: meibomian gland region segmentation, meibomian gland region segmentation, and gland boundary detection. Considering that meibomian glands exhibit a slender, strip-like distribution and may experience severe adhesion or rupture under pathological conditions, relying solely on general semantic segmentation models often fails to achieve sub-glandular level quantization accuracy. Therefore, this application designs the following structure: First, the input eyelid image is fed into an encoder for feature extraction. The encoder can use a ResNet pre-trained on ImageNet as its backbone network, extracting multi-scale features of the image through layer-by-layer downsampling. The feature map output by the encoder is then fed into a Strip Mixed Attention Module (SMAM). Considering the anisotropic geometric features of the meibomian glands, this module utilizes a strip pooling mechanism, capturing long-distance vertical and horizontal dependencies through long strip convolutional kernels in the horizontal direction (1×K) and vertical direction (K×1). This design adapts more effectively to the elongated contour of the glands than the traditional square receptive field, enhancing the model's ability to recognize slender glands, adherent glands, and broken glands.

[0051] The enhanced feature maps output by the strip-mixed attention module are fed into two branches: an attention gating module and a boundary detection branch. The boundary detection branch (BDA-Net) is specifically designed to identify narrow gaps between adjacent meibomian glands. By introducing boundary constraint features, it forces the network to learn to distinguish the edges of two closely spaced glands, outputting a boundary segmentation mask between adjacent meibomian glands. The attention gating module employs an attention-gated U-Net structure, introducing an additive activation attention gate in the skip connections between the encoder and decoder. This gating structure combines local feature maps from the contraction path and global gating signals from the decoding path, adaptively ignoring non-glandular textures in the input image, suppressing background noise features, and enhancing the response of the meibomian gland region, significantly improving segmentation accuracy. Furthermore, to address the issue of fragmented gland morphology in pathological images, some solutions incorporate VAE-GAN as a quality monitor. Variational autoencoders (VAEs) learn the manifold representation of glands, providing topological constraint information for segmentation masks. Through adversarial training (Discriminator), broken blocks or isolated noise points that do not conform to physiological characteristics are eliminated, ensuring that the segmented glands are geometrically continuous.

[0052] The feature map processed by the attention gating module is input into the decoder for upsampling and feature reconstruction, gradually restoring the spatial resolution of the image. Finally, multiple segmentation heads output the meibomian gland region segmentation mask and the meibomian plate region segmentation mask respectively.

[0053] Through the above structural design, the unified segmentation model can simultaneously output region segmentation results and boundary detection results. Among them, the strip hybrid attention module effectively adapts to the slender geometric characteristics of glands; the boundary detection branch specifically identifies the narrow gap boundaries between adjacent glands, providing key information for subsequent separation of adherent glands; the attention gating module suppresses background noise and enhances the response of gland regions, thereby improving the accuracy of segmentation.

[0054] In addition, the region segmentation mask of the smallest overall range of the meibomian gland can be output by the segmentation head to assist in the calculation of the meibomian gland atrophy rate based on the contour envelope method, and compared with the atrophy rate based on individual meibomian gland instance segmentation to evaluate the sensitivity advantage of the individualized segmentation method proposed in this application in early lesion detection compared with the traditional contour envelope method.

[0055] Therefore, this application constructs a unified segmentation model capable of simultaneously handling region segmentation and boundary detection tasks through the collaborative efforts of an encoder, a strip hybrid attention module, an attention gating module, a decoder, a boundary detection branch, and multiple segmentation heads. This model is specifically optimized for the pathological characteristics of meibomian glands—their elongated shape and tendency to adhere and break—and can accurately identify elongated, adhered, broken, and pathologically deformed meibomian gland structures, improving segmentation accuracy in complex scenarios and providing high-quality region and boundary segmentation masks for subsequent instance segmentation and refined quantitative analysis.

[0056] Optionally, the unified segmentation model further includes a highlight region detection branch and an image restoration branch; and before inputting the eyelid image into the encoder for feature extraction, it further includes: determining whether the eyelid image has reflective highlight artifacts; if not, directly inputting the eyelid image into the encoder for feature extraction; if so, performing the following restoration operation: obtaining the mask of the reflective highlight artifacts in the eyelid image through the highlight region detection branch; superimposing the mask of the reflective highlight artifacts with the eyelid image through the image restoration branch, and performing texture completion and edge reconstruction on the occluded area of ​​the reflective highlight artifacts, outputting the restored image after highlight removal, and then inputting the restored image into the encoder for feature extraction.

[0057] Specifically, in combination Figure 3As shown, the unified segmentation model of this application sets up a highlight region detection branch and an image restoration branch before the encoder to solve the common problem of reflective highlight artifacts in infrared meibomian gland images. During clinical acquisition, due to tear film reflection or the presence of droplets, highlight artifact regions often appear in infrared images. These regions can obscure the underlying meibomian gland tissue, affecting the accuracy of segmentation. Existing automated methods usually require tedious, separate processing of these reflective highlight artifacts outside the model to obtain a satisfactory segmentation result, increasing the complexity of the operation. To address this problem, this application integrates the processing of reflective highlight artifacts into the model, achieving a "one-step" solution.

[0058] The specific restoration process is as follows: Before inputting the eyelid image into the encoder for feature extraction, it is first determined whether there are reflective highlight artifacts in the image. If there are no reflective highlight artifacts, the original image is directly input into the encoder for feature extraction; if there are reflective highlight artifacts, the reflective highlight problem is converted into an image restoration problem, and the image restoration process is initiated. Image restoration refers to filling in missing, occluded, or abnormal areas in the image, so that the restored image maintains a natural and structural continuity. The image restoration process specifically includes: First, the input image is analyzed by a highlight region detection branch, which learns to identify and locate specific regions of reflective highlight artifacts in the image and outputs a mask for that region. Then, this mask is overlaid on the original image, and the image inpainting branch performs texture completion and edge reconstruction on the highlight occlusion regions indicated by the mask. The reflective highlight problem is transformed into an image inpainting problem. Given an input image I and a highlight region mask M (M=1 represents a known region, M=0 represents a highlight region to be repaired), the image inpainting branch learns the mapping function f(·) to generate the repaired image. The expression is: The image inpainting branch learns the texture features of surrounding normal glandular tissue to infer and fill in the content of the occluded area, generating a de-highlighted inpainted image. Finally, the inpainted image is input into the encoder for subsequent feature extraction and segmentation. Through the above design, this application integrates the processing of reflective highlight artifacts into the model, achieving non-destructive inpainting of image artifacts. During the model training phase, by including highlighted areas as independent labeled categories in the training data, the model can autonomously learn to recognize reflective highlight artifacts in the image and the true structural features of the surrounding glandular tissue.

[0059] Therefore, this application achieves automatic detection and repair of reflective highlight artifacts in infrared meibomian gland images by using a highlight region detection branch and an image restoration branch. This design integrates the cumbersome external artifact processing process of traditional methods into the model itself. The repaired image can be directly obtained from the original input image, and subsequent segmentation can be completed without additional external artifact processing on the acquired eyelid images, simplifying the operation process and improving diagnostic efficiency. Simultaneously, through texture completion and edge reconstruction, the glandular structure information in the occluded areas is preserved, avoiding information loss due to artifacts and providing a high-quality image foundation for subsequent accurate segmentation and quantitative analysis.

[0060] Optionally, the multidimensional morphological parameters include gland length, average width, 10th percentile width, 90th percentile width, diameter deformation index, tortuosity, proportion of meibomian gland parenchyma per unit area, meibomian gland atrophy rate, meibomian gland density, and signal intensity index; and, based on the segmentation results, extracting multidimensional morphological parameters for each independent meibomian gland includes: fitting the master curve to the instance segmentation mask of each independent meibomian gland, extracting the gland's central skeleton, and integrating to calculate the actual arc length of the central skeleton to obtain the gland length; constructing normal directions at each sampling point of the central skeleton, calculating the distance between the intersection of the normal and the two side boundaries of the gland to obtain the local width distribution, and adjusting the local width distribution accordingly. The average width is obtained by averaging the width distribution of the local area. The 10th percentile width and the 90th percentile width are obtained by performing percentile statistics on the local width distribution. The width fluctuation degree is calculated based on the local width distribution to obtain the diameter deformation index. The arc-chord ratio is calculated based on the actual arc length of the gland and the straight-line distance between the two ends of the gland to obtain the tortuosity. The percentage of meibomian gland parenchyma, the meibomian gland atrophy rate, and the meibomian gland density are calculated based on the instance segmentation mask of each independent meibomian gland and the region segmentation mask of the tarsal plate. The signal intensity index is calculated based on the average gray level inside the instance segmentation mask of each independent meibomian gland and the average gray level of the surrounding tissue region.

[0061] Specifically, after obtaining the instance segmentation mask for each independent meibomian gland and the region segmentation mask for the meibomian gland, this application employs principal curve fitting technology to extract morphological parameters. The principal curve is a nonlinear generalization of principal component analysis, finding a smooth curve passing through the center of all pixels in the gland through iterative optimization. For each instance segmentation mask of an independent meibomian gland, principal curve fitting is first performed to extract the central skeleton of the gland, and then the actual arc length of the central skeleton is calculated by integration to obtain the gland length. The gland length L is defined as the integral arc length of this curve, expressed as: Where f(s) represents the fitted principal curve, which is a curve function with arc length parameter s as the independent variable. bottom and top represent the upper and lower limits of integration, corresponding to the start and end points of the principal curve, respectively. x(s) and y(s) represent the parametric equations of the principal curve, which are the x-coordinates and y-coordinates of each point on the curve with arc length parameter ss as the independent variable.

[0062] The final gland length is then obtained by using a conversion factor of pixels per millimeter (mm / pixel).

[0063] This method has higher stability when dealing with edge burrs compared to traditional distance-transformation-based skeletonization algorithms.

[0064] In terms of width parameter extraction, instead of simply calculating the average width of the meibomian gland, the local width is measured point-by-point along the axial direction based on the central skeleton of the gland to obtain a width distribution sequence reflecting the morphological changes of the gland throughout its entire course, thereby achieving a refined quantitative analysis of the width characteristics of the meibomian gland. Preferably, the local width is obtained by: fitting a principal curve to the segmented single meibomian gland region to extract its central skeleton; constructing a normal direction at each sampling point of the skeleton and calculating the distance between the intersection of the normal and the two side boundaries of the gland as the local width value at that location; repeating the above operation for all sampling points to obtain the local width distribution of the meibomian gland. The average width is obtained by averaging the local width distribution.

[0065] Simultaneously, percentile statistics were performed on the local width distribution, introducing the 10th percentile width and the 90th percentile width. The 10th percentile width is used to characterize the representative width of thinner areas of the gland, reflecting the degree of glandular atrophy or thinning; the 90th percentile width is used to characterize the representative width of thicker areas of the gland, reflecting the degree of glandular dilation or compensatory thickening. Compared to the maximum, minimum, or average values, the percentile width index is insensitive to local noise points and abnormal boundary disturbances, and can more stably and objectively describe the distribution characteristics of meibomian gland width. Therefore, by jointly analyzing the 10th percentile width and the 90th percentile width, it is possible to simultaneously identify two different types of pathological changes in the meibomian gland: local thinning and local thickening, compensating for the inadequacy of a single average width index in fully reflecting the heterogeneity of width changes. The degree of width fluctuation is calculated based on the local width distribution to obtain the Diameter Deformation Index (DI). This index is used to quantify the dramatic fluctuations in the axial width of the gland. When the gland undergoes local expansion or narrowing due to pressure, the Diameter Deformation Index increases significantly, which helps to identify atypical MGD progression at an early stage.

[0066] In addition, the 10th percentile width and 90th percentile width can be used for longitudinal comparative analysis before and after treatment; when the 10th percentile width increases after treatment, it suggests that the narrow area of ​​the gland may be repaired; when the 90th percentile width decreases or tends to the normal range after treatment, it suggests that the abnormal expansion of the gland may be improved; together, they can serve as objective quantitative indicators for evaluating the efficacy of hot compresses, intense pulsed light therapy, meibomian gland massage, or other interventions.

[0067] Regarding the extraction of distortion parameters, this application integrates multiple models for comprehensive evaluation of distortion. The first is the arc-chord ratio model, defined as TI = Larc / Lchord-1, which is the ratio of the actual principal curve arc length Larc of the gland to the Euclidean distance Lchord between the two endpoints of the gland, minus 1. This is used to assess the length changes and morphological shortening of the gland caused by lesions. The second is the convex hull ratio model, which is the ratio of the gland area to the area of ​​the smallest circumscribed convex polygon. This is used to assess the gland's filling efficiency of the surrounding space; the higher the distortion, the lower this ratio. The third is the circumscribed rectangle height ratio model, calculated as MGtortuosity = MGperimeter / (2×H)-1, where MGperimeter is the gland perimeter and H is the height of the smallest circumscribed rectangle. This is used to accurately quantify the degree of compensatory distortion growth after gland damage.

[0068] Regarding the extraction of distribution and structural parameters, the percentage of meibomian gland parenchyma, the meibomian gland atrophy rate, and the meibomian gland density are calculated based on the instance segmentation mask of each independent meibomian gland and the region segmentation mask of the tarsal plate. The percentage of individual meibomian gland parenchyma is calculated using individualized segmentation implemented with Mask R-CNN, determining the proportion of parenchymal pixels of each independent gland to the tarsal plate region. Experimental data show that the individual analysis method has significantly better sensitivity (area difference ratio 2.12 times) in distinguishing lesions than the traditional contour envelope method (1.68 times), enabling earlier detection of localized parenchymal atrophy. The meibomian gland atrophy rate is defined as the percentage (Loss) of the area of ​​the gland's absence relative to the total tarsal area, i.e., Loss = 1 - (AreaGland / AreaEyelid), which is the core logic for achieving automated Meiboscore standard grading. Here, AreaGland represents the meibomian gland area, i.e., the total pixel area covered by the instance segmentation mask of each independent meibomian gland output by the segmentation model. When calculating the overall atrophy rate, it usually refers to the total pixel area of ​​all visible meibomian glands. AreaEyelid represents the area of ​​the tarsal plate region, which is the total area of ​​pixels covered by the tarsal plate region segmentation mask output by the segmentation model, corresponding to the complete area of ​​the tarsal bone (tarsal plate) in the eyelid.

[0069] Meibomian gland density is defined as the ratio of the total number of visible gland pixels in the entire tarsal plate area to the total number of pixels in the tarsal bone, reflecting the density of gland distribution.

[0070] In terms of optical and qualitative parameter extraction, the signal intensity index is calculated based on the average gray level within the segmentation mask of each individual meibomian gland instance and the average gray level of the surrounding tissue region. The signal intensity index is quantified based on the average gray level contrast between the glandular parenchyma and the surrounding tarsal tissue, and is used to identify "ghost glands"—glandules that, while not completely morphologically absent, often exhibit significantly diminished secretory function. This index provides an objective basis for assessing the function of the glandular parenchyma.

[0071] Therefore, this application, through the aforementioned multidimensional morphological parameter extraction method, achieves comprehensive quantification of the morphological structure (length, width, tortuosity), distribution characteristics (glandular density, atrophy rate, individual parenchymal proportion), and functional status (signal intensity index) of meibomian glands. In particular, by introducing the 10th and 90th percentile widths, refined width analysis is achieved; by introducing the diameter deformation index, quantitative monitoring of width fluctuations is realized; and by introducing the signal intensity index, indirect assessment of glandular parenchymal function is achieved. This allows for earlier and more sensitive capture of subtle pathological changes in MGD, providing a refined quantitative basis for subsequent differential status assessment.

[0072] In addition, such as Figure 3 As shown, before performing step S206, the segmentation results are post-processed to eliminate noise artifacts, fill local voids, and enhance the geometric continuity of the glandular region.

[0073] Preferably, the post-processing includes segmentation optimization based on topological constraints; the topological constraints are implemented through generative models, variational autoencoders, generative adversarial networks, or morphological rules, and are used to remove isolated noise points and discontinuous pseudo-segments that do not conform to physiological structures, while maintaining the geometric continuity of the glandular region.

[0074] Optionally, determining the meibomian gland classification label of the eyelid image includes: extracting the aspect ratio of the minimum bounding rectangle of the meibomian gland region and the vertical position coordinates of the centroid of the meibomian gland region in the image based on the region segmentation mask of the meibomian gland in the segmentation result; when the aspect ratio of the minimum bounding rectangle is greater than or equal to a preset threshold and the vertical position coordinates of the centroid are less than 0.5, determining the meibomian gland classification label as upper meibomian gland; when the aspect ratio of the minimum bounding rectangle is less than the preset threshold and the vertical position coordinates of the centroid are greater than or equal to 0.5, determining the meibomian gland classification label as lower meibomian gland.

[0075] Specifically, after obtaining the tarsal plate region segmentation mask from the segmentation results, this application further utilizes the geometric features of the mask to automatically identify whether the current eyelid image originates from the upper or lower eyelid. Because the upper and lower eyelids differ significantly in anatomical structure and imaging morphology, the tarsal plate region of the upper eyelid typically presents a relatively long and slender shape, with a larger aspect ratio (i.e., the ratio of the rectangle's length to its width) of its minimum bounding rectangle, and this region is positioned higher in the image (i.e., its centroid vertical coordinate is smaller); while the tarsal plate region of the lower eyelid is relatively short and thick, with a smaller aspect ratio of its minimum bounding rectangle, and its position is lower (i.e., its centroid vertical coordinate is larger). Based on this prior knowledge, this application achieves automatic classification through the following rules: First, based on the region segmentation mask of the meibomian gland, the aspect ratio of the minimum bounding rectangle of the region and the vertical coordinates of the region's centroid in the image are calculated (the image height is normalized to 1, and the coordinate values ​​range from 0 to 1); then, the aspect ratio of the minimum bounding rectangle is compared with a preset threshold, and the vertical coordinates of the centroid are compared with 0.5; when the aspect ratio of the minimum bounding rectangle is greater than or equal to the preset threshold and the vertical coordinates of the centroid are less than 0.5, the classification label is determined to be the upper meibomian gland; when the aspect ratio of the minimum bounding rectangle is less than the preset threshold and the vertical coordinates of the centroid are greater than or equal to 0.5, the classification label is determined to be the lower meibomian gland. The above preset threshold can be set according to the statistical results of the actual dataset, for example, set to 2.0.

[0076] Therefore, this application achieves automatic classification and recognition of upper and lower eyelid images through geometric feature analysis based on tarsal plate region segmentation masks. This automatic classification mechanism eliminates the need for manual input of eyelid type information, completing the process automatically within the workflow, simplifying operations and improving efficiency. Simultaneously, accurate upper and lower eyelid classification labels provide crucial information for subsequent differentiated status assessments of different eyelid locations (e.g., generating separate assessment reports for the upper and lower eyelids), further enhancing the accuracy and clinical reference value of the assessment results.

[0077] Optionally, generating the corresponding meibomian gland status assessment result based on the meibomian gland classification label and the multidimensional morphological parameters includes: generating meibomian gland attribution information of the eyelid image based on the meibomian gland classification label; automatically scoring or assisted grading the meibomian gland status based on the multidimensional morphological parameters; and performing longitudinal comparative analysis of the multidimensional morphological parameters at different time points for the same patient, outputting the change trend of at least one morphological parameter.

[0078] Specifically, after extracting multidimensional morphological parameters and determining the classification labels for the upper and lower eyelids, this application further generates structured meibomian gland status assessment results based on this information to serve clinical diagnosis and efficacy evaluation.

[0079] First, based on the meibomian gland classification label determined in step S206, the meibomian gland attribution information of the eyelid image is generated, that is, the evaluation report clearly indicates whether the currently analyzed eyelid image originates from the upper eyelid or the lower eyelid. Since the upper and lower eyelids differ in anatomical structure and pathological manifestations, presenting the evaluation results separately helps doctors make targeted diagnostic and treatment decisions.

[0080] Secondly, based on the aforementioned multidimensional morphological parameters, the status of the meibomian glands is automatically scored or assisted in grading. The automatic scoring includes the Meiboscore score or other grading methods based on the proportion of glandular area loss, the degree of morphological disorder, and changes in grayscale signal.

[0081] The meibomian gland atrophy rate is defined as the percentage of the area of ​​gland absence relative to the total tarsal area, calculated using the formula Loss = 1 - (AreaGland / AreaEyelid). Based on this continuous value, the system can automatically convert the meibomian gland atrophy rate into a standard Meiboscore grading (e.g., 0-3 or 0-6). For example, an atrophy rate less than 33% is graded as grade 1, 33% to 66% as grade 2, and greater than 66% as grade 3. Compared to the traditional method of manually estimating vague proportions like "about one-third" by the naked eye, this application achieves a more accurate standardized grading by calculating the percentage atrophy rate (e.g., 33.5%) based on precise pixel-level segmentation results, avoiding the subjective differences inherent in manual assessment.

[0082] Finally, a longitudinal comparative analysis of the multidimensional morphological parameters at different time points for the same patient is performed, outputting the trend of change of at least one morphological parameter. In clinical practice, for patients receiving interventions such as hot compresses, intense pulsed light (IPL) therapy, and meibomian gland massage, doctors typically need to collect eyelid images at multiple time points before, during, and after treatment to assess the treatment effect. This application supports the analysis of images from the same patient at different time points, and the longitudinal comparison of multidimensional morphological parameters (such as gland length, average width, 10th percentile width, 90th percentile width, tortuosity, atrophy rate, signal intensity index, etc.) at each time point, outputting the trend of change of each parameter. For example, when the 10th percentile width increases after treatment, it suggests that the narrow area of ​​the gland may be repaired; when the 90th percentile width decreases or approaches the normal range after treatment, it suggests that the abnormal expansion of the gland may be improved; when the signal intensity index increases, it suggests that the parenchymal function of the gland may have recovered. The trends of these quantitative indicators provide clinicians with objective and accurate evidence for efficacy evaluation.

[0083] Spearman correlation analysis showed that the automatically extracted morphological indicators were highly correlated with clinical MGD scores. Specifically, tortuosity (TI) and diameter deformity index (DI) achieved AUCs of 0.96 and 0.82, respectively, in distinguishing healthy individuals from early-stage MGD patients. Furthermore, the model demonstrated a sensitivity of 84.4% and a specificity of 71.7% in identifying "ghost glands," providing an indirect means of assessing glandular parenchymal function.

[0084] Therefore, this application achieves a comprehensive and structured assessment of meibomian gland status by generating meibomian gland attribution information, Meiboscore scores, and longitudinal comparative analysis results. Specifically, the Meiboscore score automatically generated based on meibomian gland atrophy rate overcomes the experience dependence and subjective variability of traditional manual scoring; and the longitudinal comparative analysis based on multi-timepoint images enables objective quantitative monitoring of treatment efficacy. These assessment results provide comprehensive and objective data support for the clinical diagnosis, disease grading, treatment plan formulation, and efficacy evaluation of MGD, effectively improving clinical work efficiency and diagnostic consistency.

[0085] Preferably, the method can also be fused with non-contact tear film breakup time, tear lipid layer thickness, or dry eye-related examination parameters to form a multimodal joint evaluation result.

[0086] Preferably, in cases of severe meibomian gland atrophy, the glandular tissue occupies only a small portion of the tarsal bone area, leading to a severe class imbalance problem. Traditional cross-entropy loss (BCE) tends to predict the background in such situations. To optimize the loss function for different disease stages, this approach employs a combined loss function during model training. This combined loss function includes at least Dice_Loss and Focal_Loss, and for scenarios requiring fine separation of adjacent glands, it further includes cross-entropy loss (BCE_Loss) to reduce the distance between pixel features within the same gland and increase the distance between pixel features between different glands, thereby improving the ability to separate adherent glands. Therefore, the expression for the combined loss function is as follows: Where α, β, and γ are preset weights.

[0087] To further verify the effectiveness and practicality of the technical solution of this application, the following provides specific experimental verification data, alternative technical solutions, and a discussion of clinical application value.

[0088] The following experimental environment was used in the model training and validation phases of this application: Hardware environment: The compute node is equipped with an NVIDIA GeForce RTX 3090 GPU (24GB VRAM), an Intel Core i9-10900K CPU, and 64GB RAM.

[0089] Software framework: Based on the PyTorch deep learning framework, using CUDA 11.1 for hardware acceleration.

[0090] Dataset size: 1,087 infrared meibomian gland images were collected from dry eye clinics, including 629 images of the upper eyelid and 458 images of the lower eyelid. The dataset was randomly divided into training, validation, and test sets in a 70%:20%:10% ratio.

[0091] Preprocessing parameters: All original images were uniformly cropped to 420×890 pixels. A bilateral filter (kernel size 3) was applied for smoothing. Data augmentation was performed using a 16x amplification method (including color dithering, random lighting, and horizontal flipping).

[0092] Taking the SBD-MTLNet model as an example, the model training process is as follows: Parameter initialization: ResNet18 pre-trained on ImageNet is used as the Encoder weights to accelerate model convergence.

[0093] Training strategy: The Adam optimizer is used, with an initial learning rate of 0.0001, which decays to 0.5 every 10 epochs. The batch size is set to 5.

[0094] Loss function configuration: set α=0.5, β=0.25, γ=0.25.

[0095] Training period: The total number of training iterations was 30 epochs, and the entire training process took about 8 hours and 24 minutes.

[0096] After obtaining the pixel mask of the test set image, quantization analysis is performed according to the following logic: Physical scale mapping: First, a standard calibration grid image is captured, and the digital resolution coefficient of the system is calculated (e.g., R_res=0.015mm / pixel).

[0097] Example of length (L) calculation: For a segmented gland, the skeleton is extracted using master curve fitting. This skeleton contains a total of 430 pixels. Therefore, its actual physical length is 430 × 0.015 = 6.45 mm.

[0098] Example of Twist Index (TI): The actual arc length of a gland is 6.45 mm, and the Euclidean distance between its two endpoints is 5.70 mm. According to the formula TI = (6.45 / 5.70) - 1 = 0.13.

[0099] This application evaluated the segmentation performance of the tarsal plate region and meibomian glands on a test set, using metrics including Dice coefficient, recall, and precision. The specific evaluation results are as follows: The results of tarsal plate region segmentation were as follows: Dice coefficient was 0.928384, recall was 0.902173, and precision was 0.964809.

[0100] Meibomian gland segmentation results: Dice coefficient was 0.739411, recall was 0.778219, and precision was 0.72534.

[0101] The above evaluation results show that the multi-task learning network proposed in this application has achieved good segmentation performance in both the meibomian region and meibomian gland segmentation tasks, which can meet the needs of clinical quantitative analysis.

[0102] The clinical value of this plan lies in its ability to improve the level of ophthalmological clinical practice from multiple dimensions, specifically: (1) Optimization of clinical diagnostic process: Traditional manual assessment of each image takes an average of 5 to 8 minutes and is easily affected by fatigue. The inference time of the automated solution is usually between 0.32 and 0.5 seconds, which can realize real-time screening report generation. This makes it possible to integrate meibomian gland assessment in large-scale physical examination screening.

[0103] (2) Longitudinal monitoring and efficacy evaluation: For patients receiving meibomian gland heat therapy, IPL (intense pulsed light) therapy, or meibomian gland massage, subtle morphological changes (such as slight reduction in width and increase in signal intensity) are important criteria for judging efficacy. Quantitative indicators provide continuous variables that can accurately monitor changes in the course of the disease over time through statistical methods, providing closed-loop data support for individualized precision treatment.

[0104] Therefore, this solution can shorten the traditional 5-8 minute manual assessment process to less than 0.5 seconds and provide quantitative indicators that surpass visual observation. These quantitative indicators can accurately monitor subtle changes in efficacy after IPL treatment or interventions such as warm compresses, opening up a new data-driven path for the individualized and precise management of dry eye syndrome. With the evolution of algorithms and the fusion of multimodal data, this technology is expected to become a standard feature of intelligent ophthalmological diagnosis and treatment.

[0105] In summary, this application, through the combination of the above series of technical solutions, achieves automated segmentation and multidimensional morphological quantification of meibomian gland images, providing an efficient and accurate intelligent tool for the clinical diagnosis, disease assessment, and efficacy monitoring of meibomian gland dysfunction.

[0106] To further expand the scope of protection of this application and address the challenges of different hardware environments or data distributions, this application provides the following optional alternative technical solutions: (1) Alternatives to image enhancement: In addition to the CLAHE method, this approach can employ image enhancement models based on generative adversarial networks (GANs), such as mirror-symmetric GANs, to remove specular reflection artifacts caused by tear film reflection. This alternative learns the feature distribution of high-quality images through adversarial training, restoring glandular details while suppressing noise. Furthermore, a frequency-aware adjustment module (FAM) can be used to decompose the image into different frequency bands using Fourier transform, enhancing low-frequency glandular responses and suppressing mid-to-high-frequency noise backgrounds.

[0107] (2) Alternatives to the backbone network and attention mechanism: The main solution of this application adopts SBD-MTLNet. An alternative is to use a Transformer architecture (such as TransUnet) as the backbone network. Transformer uses a self-attention mechanism to capture the global dependencies of the image, which can better handle the long-span tortuous morphology of meibomian glands. In addition, in the attention module, an additive activation attention gate can be used instead of the strip pooling module, which can locate the gland region by fusing the feature maps of skip connections and the gating signal.

[0108] (3) Alternative solution for segmenting adhering glands: In addition to the boundary detection auxiliary network (BDA-Net), the alternative solution for the problem of adhering adjacent glands is to use traditional computer vision morphology processing. The specific steps are as follows: First, perform distance transformation on the initially segmented binary mask to calculate the distance of each pixel to the nearest background. Then, apply the watershed algorithm to the distance map to perform region growing, thereby achieving forced separation of adhering glands at the narrowest physical gap.

[0109] (3) Alternative to labeling and supervision strategies: In the absence of high-precision pixel-level labeled data, this application can adopt a weak supervision alternative, namely a training strategy based on scribble annotation. This approach introduces temporal ensemble prediction and transformation equivariance constraint, which can achieve near-fully supervised segmentation performance at the cost of reducing labeling costs by 80%.

[0110] (4) Alternative solutions for morphological index measurement: In the morphological quantification stage, in addition to master curve fitting, an alternative solution can be a method based on binary skeletonization combined with principal component analysis (PCA). This method extracts the central axis by refining the segmented region and uses PCA to capture the principal geometric axis of the gland, thereby calculating the length and width. For the degree of tortuosity, an information entropy model can be used to quantify the tortuosity index by calculating the degree of disorder in the spatial distribution of the target curve relative to the standard straight line.

[0111] (5) Alternative to the combined loss function: This application may use Focal Tversky Loss (FTL) to replace the original combined loss. FTL adjusts the penalty weights for false positives and false negatives by introducing α and β parameters, and adds a γ focusing index on the basis of the Tversky index, so that the model still has extremely high detection sensitivity in severe pathological images with very small glandular areas.

[0112] In summary, this application proposes a deep learning-based method for meibomian gland image segmentation and morphological quantization, with four core innovations: First, it achieves comprehensive coverage and individualized instance segmentation of the meibomian glands in both the upper and lower eyelids. By constructing a standardized dataset that includes both upper and lower eyelid images, and designing a multi-task learning network that includes a unified segmentation model and an adaptive interactive information fusion module, it can not only output the segmentation mask of the meibomian gland region, but also the instance segmentation mask of each individual meibomian gland, effectively overcoming the shortcomings of existing technologies that focus mainly on the upper eyelid and can only output the overall outline.

[0113] Secondly, comprehensive quantification of multidimensional morphological parameters was achieved. Based on the instance segmentation mask of each individual meibomian gland, multidimensional morphological parameters, including gland length, average width, 10th and 90th percentile width, diameter deformation index, tortuosity, individual meibomian gland parenchyma ratio, meibomian gland atrophy rate, meibomian gland density, and signal intensity index, were extracted through methods such as master curve fitting, local width distribution analysis, arc-chord ratio calculation, and gray-scale contrast analysis. This enabled a comprehensive quantitative assessment of the morphological structure, distribution characteristics, and functional status of the meibomian glands. Furthermore, the introduction of the 10th and 90th percentile width indicators not only reflects glandular atrophy (thinning) but also monitors compensatory expansion (thickening), providing a highly valuable longitudinal observation indicator for evaluating glandular repair after treatment.

[0114] Third, it achieves "one-step" processing of reflective highlight artifacts. By setting highlight region detection and image inpainting branches in the unified segmentation model, the cumbersome external artifact processing process in traditional methods is integrated into the model. The original image can be directly input to obtain the repaired image and complete subsequent segmentation without the need for additional preprocessing of the acquired eyelid images, simplifying the operation process and improving diagnostic efficiency.

[0115] Fourth, it enables structured status assessment and longitudinal monitoring. Eyelid classification information is generated based on upper and lower eyelid labels, and Meiboscore scores are automatically generated based on meibomian gland atrophy rates. Furthermore, longitudinal comparative analysis of multidimensional morphological parameters at different time points for the same patient is performed, outputting the trends of each parameter's changes. This provides comprehensive and objective data support for clinical diagnosis, disease grading, and efficacy evaluation.

[0116] Compared with the prior art, this application has the following beneficial effects: (1) By establishing a standardized infrared meibomian gland image preprocessing process, this application can effectively improve the problems of insufficient image contrast, large noise interference and obvious reflection artifacts, thereby improving the stability and applicability of the subsequent segmentation model.

[0117] (2) This application, through the joint design of multi-task learning network, boundary detection branch and attention mechanism, can more accurately identify slender, adhered, broken and pathologically deformed meibomian gland structures, and improve the segmentation accuracy in complex scenes.

[0118] (3) This application can effectively alleviate the class imbalance problem in severe atrophy cases by combining loss function and topological constraint post-processing mechanism, reduce false positive and false negative results, and improve the physiological rationality of segmentation results.

[0119] (4) Based on the segmentation, this application further automatically extracts multi-dimensional quantitative indicators such as length, width, tortuosity, atrophy rate, density and signal intensity, realizing objective analysis of the morphological structure and degree of degeneration of meibomian glands, overcoming the defects of manual observation that rely on experience and lacks quantification.

[0120] (5) This application can automatically generate Meiboscore scores and morphological analysis reports. It has a fast reasoning speed and can be used for outpatient screening, large-scale physical examinations and efficacy follow-up, which is conducive to improving clinical work efficiency and diagnostic consistency.

[0121] (6) This application can be extended to different devices, different centers and multimodal data fusion scenarios, and has good clinical application prospects and promotion value.

[0122] In addition, refer to Figure 1 As shown, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0123] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0124] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application. Example 2

[0125] Figure 9 A deep learning-based meibomian gland image segmentation and morphological quantization apparatus according to this embodiment is shown, which corresponds to the method described in Embodiment 1. Reference Figure 9As shown, the device includes: an image acquisition module 910, used to acquire eyelid images collected by an infrared imaging device, wherein the eyelid images are either upper eyelid images or lower eyelid images; an image segmentation module 920, used to input the eyelid images into a pre-trained multi-task learning network and output segmentation results; the segmentation results include instance segmentation masks for each independent meibomian gland and region segmentation masks for the tarsal plate; a parameter extraction module 930, used to extract multidimensional morphological parameters for each independent meibomian gland based on the segmentation results and determine the meibomian gland classification label of the eyelid image; the multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional state of the meibomian gland; the meibomian gland classification label is either upper meibomian gland or lower meibomian gland; and a state evaluation module 940, used to generate corresponding meibomian gland state evaluation results based on the meibomian gland classification label and the multidimensional morphological parameters.

[0126] It should be noted that the meibomian gland image segmentation and morphological quantization device based on deep learning provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0127] Therefore, according to this embodiment, an end-to-end deep learning model is used to achieve comprehensive coverage of the meibomian glands of the upper and lower eyelids, individualized instance segmentation, and automated extraction and classification evaluation of multidimensional morphological parameters. This solves the technical problems of incomplete evaluation scope, lack of individual gland instance segmentation capability, single morphological parameters, and cumbersome external processing of image artifacts in existing automated methods. Example 3

[0128] Figure 10 A deep learning-based meibomian gland image segmentation and morphological quantization apparatus according to this embodiment is shown, which corresponds to the method described in Embodiment 1. Reference Figure 10 As shown, the device includes: a processor 1010; and a memory 1020 connected to the processor 1010, used to provide the processor 1010 with instructions to process the following steps: acquiring an eyelid image captured by an infrared imaging device, wherein the eyelid image is either an upper eyelid image or a lower eyelid image; inputting the eyelid image into a pre-trained multi-task learning network and outputting a segmentation result; the segmentation result includes an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; based on the segmentation result, extracting multidimensional morphological parameters for each independent meibomian gland and determining the meibomian gland classification label for the eyelid image; the multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional state of the meibomian gland; the meibomian gland classification label is either an upper meibomian gland or a lower meibomian gland; and generating a corresponding meibomian gland state assessment result based on the meibomian gland classification label and the multidimensional morphological parameters.

[0129] It should be noted that the meibomian gland image segmentation and morphological quantization device based on deep learning provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0130] Therefore, according to this embodiment, an end-to-end deep learning model is used to achieve comprehensive coverage of the meibomian glands of the upper and lower eyelids, individualized instance segmentation, and automated extraction and classification evaluation of multidimensional morphological parameters. This solves the technical problems of incomplete evaluation scope, lack of individual gland instance segmentation capability, single morphological parameters, and cumbersome external processing of image artifacts in existing automated methods.

[0131] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0132] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0133] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0134] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0135] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0136] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0137] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A deep learning-based image segmentation and morphological quantization method for meibomian glands, characterized in that, include: Acquire eyelid images captured by an infrared imaging device, wherein the eyelid images are either upper eyelid images or lower eyelid images; The eyelid image is input into a pre-trained multi-task learning network, which outputs segmentation results; the segmentation results include an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; Based on the segmentation results, multidimensional morphological parameters of each independent meibomian gland are extracted, and the meibomian gland classification label of the eyelid image is determined. The multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional status of the meibomian glands; the meibomian gland classification label is upper meibomian gland or lower meibomian gland. Based on the meibomian gland classification labels and the multidimensional morphological parameters, corresponding meibomian gland status assessment results are generated.

2. The method according to claim 1, characterized in that, The multi-task learning network includes a unified segmentation model and an adaptive interactive information fusion module; and the eyelid image is input into the pre-trained multi-task learning network to output segmentation results, including: The eyelid image is input into the unified segmentation model, which outputs the meibomian gland region segmentation mask, the meibomian plate region segmentation mask, and the boundary segmentation mask between adjacent meibomian glands. The region segmentation mask of the meibomian gland and the boundary segmentation mask between adjacent meibomian glands are input into the adaptive interactive information fusion module for feature fusion, and the instance segmentation mask of each independent meibomian gland is output.

3. The method according to claim 2, characterized in that, The unified segmentation model includes an encoder, a strip hybrid attention module, an attention gating module, a decoder, a boundary detection branch, and multiple segmentation heads; Furthermore, the eyelid image is input into the unified segmentation model, which outputs a region segmentation mask for the meibomian glands, a region segmentation mask for the tarsal plate, and a boundary segmentation mask between adjacent meibomian glands, including: The eyelid image is input into the encoder for feature extraction, and the features output by the encoder are input into the strip hybrid attention module to extract long-distance dependent features through horizontal and vertical strip convolution kernels; The features output by the strip hybrid attention module are respectively input into the attention gating module and the boundary detection branch; The boundary detection branch identifies the narrow gap boundary between adjacent meibomian glands and outputs the boundary segmentation mask between adjacent meibomian glands. The features output by the attention gating module are input into the decoder for upsampling and feature reconstruction, and the features output by the decoder are input into the multiple segmentation heads to output the meibomian gland region segmentation mask and the meibomian plate region segmentation mask, respectively.

4. The method according to claim 3, characterized in that, The unified segmentation model also includes a highlight region detection branch and an image restoration branch; Furthermore, before inputting the eyelid image into the encoder for feature extraction, the process also includes: Determine whether the eyelid image contains reflective highlight artifacts; If the eyelid does not exist, the eyelid image is directly input into the encoder for feature extraction; if it exists, the following repair operation is performed: The mask of the reflective highlight artifact in the eyelid image is obtained through the highlight area detection branch; The image restoration branch overlays the mask of the reflective highlight artifact with the eyelid image, performs texture completion and edge reconstruction on the occluded area of ​​the reflective highlight artifact, and outputs the restored image after highlight removal. Then, the restored image is input into the encoder for feature extraction.

5. The method according to claim 1, characterized in that, The multidimensional morphological parameters include gland length, average width, 10th percentile width, 90th percentile width, diameter deformation index, tortuosity, individual meibomian gland parenchyma ratio, meibomian gland atrophy rate, meibomian gland density, and signal intensity index. Furthermore, based on the segmentation results, the multidimensional morphological parameters of each individual meibomian gland are extracted, including: For each individual meibomian gland instance segmentation mask, a master curve is fitted to extract the gland's central skeleton, and the actual arc length of the central skeleton is calculated by integration to obtain the gland length. At each sampling point of the central skeleton, a normal direction is constructed, and the distance between the intersection of the normal and the two sides of the gland is calculated to obtain the local width distribution. The average width is obtained by taking the average value of the local width distribution, and the 10th percentile width and the 90th percentile width are obtained by performing percentile statistics on the local width distribution. The degree of width fluctuation is calculated based on the local width distribution to obtain the diameter deformation index; The degree of torsion is obtained by calculating the arc-chord ratio based on the actual arc length of the gland and the straight-line distance between the two ends of the gland; Based on the instance segmentation mask of each individual meibomian gland and the region segmentation mask of the tarsal plate, the percentage of meibomian gland parenchyma, the meibomian gland atrophy rate, and the meibomian gland density of the individual are calculated. The signal intensity index is calculated based on the average gray level inside the segmentation mask of each individual meibomian gland instance and the average gray level of the surrounding tissue region.

6. The method according to claim 1, characterized in that, The determination of the meibomian gland classification label of the eyelid image includes: Based on the region segmentation mask of the tarsal plate in the segmentation result, the aspect ratio of the minimum bounding rectangle of the tarsal plate region and the vertical position coordinates of the centroid of the tarsal plate region in the image are extracted. When the aspect ratio of the minimum bounding rectangle is greater than or equal to a preset threshold and the vertical position coordinate of the centroid is less than 0.5, the meibomian gland classification label is determined to be upper meibomian gland. When the length and width of the minimum bounding rectangle are less than a preset threshold and the vertical position coordinate of the centroid is greater than or equal to 0.5, the meibomian gland classification label is determined to be the lower meibomian gland.

7. The method according to claim 1, characterized in that, The step of generating corresponding meibomian gland status assessment results based on the meibomian gland classification labels and the multidimensional morphological parameters includes: Based on the meibomian gland classification labels, generate meibomian gland attribution information for the eyelid image; Based on the aforementioned multidimensional morphological parameters, the status of the meibomian glands is automatically scored or assisted in grading. A longitudinal comparative analysis of the multidimensional morphological parameters of the same patient at different time points was performed, and the changing trend of at least one morphological parameter was output.

8. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 7 is performed by a processor.

9. A deep learning-based device for meibomian gland image segmentation and morphological quantization, characterized in that, include: The image acquisition module is used to acquire eyelid images captured by the infrared imaging device, wherein the eyelid images are either upper eyelid images or lower eyelid images; The image segmentation module is used to input the eyelid image into a pre-trained multi-task learning network and output the segmentation result; the segmentation result includes an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the meibomian gland; The parameter extraction module is used to extract multidimensional morphological parameters of each independent meibomian gland based on the segmentation results, and determine the meibomian gland classification label of the eyelid image. The multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional status of the meibomian glands; the meibomian gland classification label is upper meibomian gland or lower meibomian gland. The status assessment module is used to generate corresponding meibomian gland status assessment results based on the meibomian gland classification labels and the multidimensional morphological parameters.

10. A deep learning-based device for meibomian gland image segmentation and morphological quantization, characterized in that, include: processor; A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: Acquire eyelid images captured by an infrared imaging device, wherein the eyelid images are either upper eyelid images or lower eyelid images; The eyelid image is input into a pre-trained multi-task learning network, which outputs segmentation results; the segmentation results include an instance segmentation mask for each independent meibomian gland and a region segmentation mask for the tarsal plate; Based on the segmentation results, multidimensional morphological parameters of each independent meibomian gland are extracted, and the meibomian gland classification label of the eyelid image is determined. The multidimensional morphological parameters are used to characterize the morphological structure, distribution characteristics, and functional status of the meibomian glands; the meibomian gland classification label is upper meibomian gland or lower meibomian gland. Based on the meibomian gland classification labels and the multidimensional morphological parameters, corresponding meibomian gland status assessment results are generated.

Images