Detection of Meibomian Gland Dysfunction by in vivo Confocal Microscopy Based on Deep Convolutional Neural Network

Detection of Meibomian Gland Dysfunction by in vivo Confocal Microscopy Based on Deep Convolutional Neural Network | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Detection of Meibomian Gland Dysfunction by in vivo Confocal Microscopy Based on Deep Convolutional Neural Network Yi Shao, Yichen Yang, Hui Zhao, Wen-Qing Shi, Xu-Lin Liao, Ting Su, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-936418/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In vivo confocal microscopy（IVCM） is a real-time in vivo high-resolution and non-invasive imaging method that allows observation of morphological changes in the lid gland at the cellular level. We use IVCM to observe the meibomian glands and divide the 12,630 pictures obtained into six groups (normal group, normal with meibomian gland opening group, meibomian gland atrophy group, meibomian gland atrophy with obstruction group, meibomian gland obstruction group and meibomian gland obstruction with opening group), randomly select 70% of the pictures and use the ResNet34 deep learning network model for training, and use the remaining 30% of the pictures and another 12889 pictures collected from the top three hospitals as internal and external validation sets to verify the performance of the model. The results show that the model validation set recognition constructed using the deep learning method has achieved good performance, and the obtained AUROCs are all greater than 0.95. This model provides the possibility for future automatic classification and diagnosis of meibomian gland dysfuncton(MGD),and can be used for MGD related clinical auxiliary diagnosis and screening of diseases. Computational Biology Bioinformatics meibomian glands in vivo confocal microscopy deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Meibomian glands are the largest sebaceous glands in the human body. They are distributed in the upper and lower meibomian layers, with 30–40 on the upper eyelid and 20–30 on the lower eyelid and are perpendicular to the eyelid margin. The meibomian gland can synthesize, store, and secrete lipids, and these secreted lipids constitute the lipid layer of the tear film. Meibomian gland dysfunction (MGD) is a chronic, non-specific inflammation characterized by meibomian gland duct obstruction or abnormal meibomian gland secretions. In MGD, the glandular lipid secretion is impaired, the ocular surface cannot maintain the stability of the tear film, the tear film evaporates rapidly, and the tear osmotic pressure increases, which can lead to dry eye disease (DED) 1,2 . With increasing age, the acinar epithelial cells of the meibomian gland shrink, leading to large-scale atrophy of the meibomian glands. As these glands are non-renewable, MGD reduces lipid secretion 3 . With an increase in the aging population, environmental pollution, and widespread use of video terminals, the incidence of MGD has gradually increased and become a disease of global proportions. The clinical symptoms of MGD are similar to those of DED and include dry eyes, eye irritation, blurred vision, and increased secretions. As the symptoms of MGD are not specific, it is easy to be missed, misdiagnosed, and mistreated. Therefore, objective physical signs can provide a reference for diagnosis, including the loss of meibomian glands, abnormal secretion of meibomian glands, and changes in the morphology of the eyelid margin. In vivo confocal microscopy (IVCM) is a new type of high-resolution, non-invasive ocular surface imaging detector. The laser focusing principle of IVCM is used to scan each point of the detected layer in the confocal plane. In IVCM, the formation of ocular surface tissues can be observed at multiple levels, and the cell level of tissue morphology and changes can be photographed in real time 4 . Hence, IVCM has real-time, non-invasive, three-dimensional advantages. In recent years, as MGD has received increasing attention, there have been an increasing number of researches on the application of IVCM in the meibomian glands 5 . IVCM can facilitate clear visualization of the characteristics of the meibomian glands, such as acinus, acinar stroma, gland secretions, periadenositis cells, and meibomian gland openings and classify them for standardized treatment. Ibrahim et al. 6 obtained IVCM images and analyzed the density of meibomian acinar units and the maximum and minimum diameters of the meibomian acinars; they concluded that IVCM has a high level of specificity and sensitivity in the diagnosis of MGD. However, because of the limitations of clinical manpower and material resources, the diagnosis of MGD-related eye diseases still has limitations at present, such as the limited surgical skills of junior doctors and the huge workload of manual classification, which are not conducive to the development of individualized disease treatment and chronic disease management. Recent advancements in the field of artificial intelligence (AI) have led to its increased application in disease assessment and diagnosis; further, AI has also become a research focus for clinical professionals. Analyzing MGD IVCM images data with high-performance AI algorithm can help distinguish different types of patients, greatly increase the work efficiency of clinicians, and save time by formulating different treatment plans for patients, in order to reduce the social and medical burden and aid with the new trend of personalized and refined MGD diagnosis and treatment. In recent years, the combined deep convolutional neural network (DCNN) model has been applied to the diagnosis of various medical diseases 7,8 . We propose a method that combines the DCNN algorithm and IVCM diagnosis for early diagnosis and classification of MGD. Methods Study Subjects The Institutional Review Board of the Ethics Committee of the First Affiliated Hospital of Nanchang University approved this retrospective study, and all methods complied with the tenets of the Helsinki Declaration. This was a single-center, clinical study. From January 2018 to June 2021, we extracted 12,630 IVCM images from the IVCM database of three hospitals, namely the First Affiliated Hospital of Nanchang University, Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine, and Shanghai General Hospital. Three ophthalmologists checked the quality of the images and performed clinical diagnosis and classification based on the 2017 Chinese Expert Consensus on the Diagnosis and Treatment of MGD 9 . The extracted data included the following: normal group (n=2896), normal with meibomian gland opening (NMGO) group (n=830), meibomian gland atrophy (MGA) group (n=3585), meibomian gland atrophy with obstruction (MGAO) group (n=1745), meibomian gland obstruction (MGO) group (n=3086), and meibomian gland obstruction with opening (MGOO) group (n=488). The six types of IVCM classification pictures can be seen in Figure 1. According to random grouping, 70% (n=8,841) and 30% (n=3,789) images were allocated to the training and internal validation sets, respectively. Simultaneously, we used 12,289 images from Nanchang University Affiliated Eye Hospital and Shanghai AIER Eye Hospital as an external verification set to verify the performance of the model. The DCNN classification model that maps the input features (such as image pixels) to the corresponding output labels (MGD of different classifications) was used to train the deep learning algorithm. We used the ResNet34 network model for classification and initial training(120 rounds of training) and developed and trained the DCNN model on the training set. We enhanced the images from the training data set using brightness, gamma correction, histogram equalization, noise addition, and inversion. The process flow chart of this research is shown in Figure 2. Classification Standard Normal group: meibomian glands and ducts are normal. NMGO group: full meibomian glands with opening. MGA group: tire-like meibomian gland epithelial cells are absent, the acinar wall disappears, becomes smaller, or the fiber cord changes. MGO group: enlarged acinus and blocked meibum in the acinus. MGAO group: shows the characteristics of both the atrophy group and obstruction group. MGOO group: normal meibomian gland opening with highly dilated meibomian gland secretory tube and blocked meibomian gland acinus. In Vivo Confocal Microscope Image Acquisition The IVCM (HRT-3, RCM module, Hidelberg Engineering, Germany) parameters were as follows: laser wavelength, 670 nm; observation field, 400 um*400 um; resolution, 384 pixels * 384 pixels; magnification, 800 times; and axis resolution, is 1 um. Use 0.5% proparacaine eye drops (American Alcon company) for topical anesthesia,and apply Carbomer Eye Gel (Shandong Bausch & Lomb Freda Pharmaceutical Co., Ltd.) on the surface of the IVCM contact lens. The disposable corneal contact cover is attached to the contact lens. To make the image clearer, the outer layer of the contact sleeve is evenly smeared with Carbomer Eye Gel. The examinee’s chin and forehead are fixed at the corresponding positions of the IVCM, and the eyelid of the examinee’s eye is turned over. The examiner fixes the eyelid with one hand to keep it in a state of eversion, and pushes the IVCM lens forward with the other hand to make it touch the edge of the examinee’s eyelid, and adjusts the scan depth, starting from the eyelid margin and gradually scanning the meibomian glands in the fornix, from the nasal side of the lesion to the temporal side, while taking IVCM images. Model Building Convolutional neural network (CNN) is a multi-layer neural network with fault tolerance and is easy to train and optimize 10 . The residual network (ResNet) is a type of CNN, the network is connected by adding one operation channel in every two layers to form a residual block to reach a jump in the whole network, avoiding the gradient descent problem and network degradation problem caused by the multiplication of very small parameters in each layer 11 . In this study, we chose a RESNET network (ResNet34) composed of 34 convolution layers to classify the MGD images. Image features were extracted using sixty-four 7×7 convolution kernels, and then a 3×3 maximum down-sampling operation with a stride of 2 was performed. Then, input the result into four residual groups with three, four, six, and three residual units, where each residual unit has two 3×3 convolutional layers with a stride of 2. After each convolutional layer, the linear rectification function (ReLU) is used as the activation function. It is mapped to the probability of each category by the fully connected layer and the Softmax activation function. In this experiment, a desktop computer was used as the experimental environment: the Central Processing Unit (CPU) is Ryzen5 1500X (Advanced Micro Devices, Inc.), the memory is 16 G, the graphic processing unit (GPU) is Nvidia GeForce GTX1060, and the graphics card memory is 6 G. We used the PyTorch (https://py-torch.org) Python framework based on the Windows 10 system to develop and train the MGD image recognition model. First, we used the pre-trained generation weight initialization network on the ImageNet (http://pytorch.org/docs/stable/torchvision/models.html) dataset, and then transferred the model network to MGD. On the data set, the modeling process is divided into three parts: image preprocessing, image feature extraction and classification of MGD IVCM, and active learning. The optimizer is Adam. The model training parameters are as follows: batch size=16, epoch=120, and learning rate=0.1; the other parameters are default. Model Evaluation and Statistical Analysis After the MGD IVCM image recognition model is trained, the performance of the trained model needs to be evaluated, and the results of classification by three ophthalmologists from the First Affiliated Hospital of Nanchang University were used as the gold standard to calculate the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, negative predictive value (NPV), Kappa coefficient, and F1 scores. Four statistical parameters are involved: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). The calculation formula is as follows: accuracy=(TP+TN)/(TP+FP+FN+TN), specificity=TN/N (negative), precision (P)=TP/(TP+FP), sensitivity=TP/P, and NPV=TN/(TN+FN). The ROC curve was plotted with specificity as the abscissa and sensitivity as the ordinate; the larger the AUROC value, the better the classification performance of the model. The Kappa test was used to compare the consistency of the results of the artificial intelligence diagnosis group and the expert group and to evaluate the accuracy of the multi-classification model. All the above indicators were obtained using the R software (version 6.078). Results Data Set Characteristics A total of 24,919 images from five hospitals were used to train and verify the performance of the DCNN. From the First Affiliated Hospital of Nanchang University, Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine, and Shanghai General Hospital, the training set consisted of 8841 images (1993 images in the normal group, 578 images in the NMGO group, 2500 images in the MGA group, 1228 images in the MGAO group, 2191 images in the MGO group, and 351 images in the MGOO group). Additionally, 3789 internal validation sets (903 in the normal group, 252 in the NMGO group, 1085 in the MGA group, 517 in the MGAO group, 895 in the MGO group, and 137 in the MGOO group), and a total of 12289 images were collected from the external validation set of Nanchang University Affiliated Eye Hospital and Shanghai AIER Eye Hospital(2912 in the normal group, 838 in the NMGO group, 3493 in the MGA group, 1693 in the MGAO group, 2866 in the MGO group, and 487 in the MGOO group). The total number, training set, internal and external validation set cohorts of the six types of MGD are summarized in Table 1. Performance of Deep Learning Algorithms We used the DCNN algorithm ResNet34 to train the classification model on the same training data set. Based on the 30% of randomly selected pictures for internal verification, the accuracy, sensitivity, specificity, precision, and AUROC parameters of the classification were calculated. The accuracy of the internal validation set test of the six groups of pictures ranged from 88.1% (95%CI: 87.1–89.2) to 93.9% (95%CI: 93.1–94.7); sensitivity, from 87.4% (95%CI: 85.0–89.5) to 93.4% (95%CI: 85.0–89.5); specificity, from 86.2% (95%CI: 84.9–87.5) to 94.3% (95%CI: 93.5–95.0); precision, from 33.8% (95%CI: 28.9–39.0) to 73.1% (95%CI: 70.7–75.5); AUROC, from 0.947 (95%CI: 0.939–0.955) to 0.971 (95%CI: 0.965–0.977). The accuracy of the six-category external validation set to verify the classification model ranged from 88.1% (95%CI: 87.5–88.6) to 93.9% (95%CI: 93.4–94.3); sensitivity, from 82.5% (95%CI: 78.9–85.8) to 94.0% (95%CI: 93.1–94.8); specificity, from 86.3% (95%CI: 85.6–87.0) to 94.3% (95%CI: 93.9–94.7); and AUROC, from 0.951 (95%CI: 0.941–0.961) to 0.971 (95%CI: 0.966–0.976). The built DCNN model internal and external validation set had high negative predictive value, Kappa coefficient, and F1 score value. The internal validation set test results are shown in the following table (Table 2). The test results of the external validation set are shown in Table 3. The ROC curve of the six classification validation sets is shown in Figures 3 and 4. Discussion In this study, we aimed to evaluate the performance of convolutional neural networks for detecting MGD using a large number of confocal microscopy images with different classifications. The results show that the combination of the DCNN model and confocal microscopy images that we tested in the validation set can finely classify MGD patients in a realistic setting with high accuracy, specificity, and AUROC values, compared to manual classification by ophthalmologists. In recent years, with the rapid development of AI, its application in clinical practice has become increasingly widespread 12,13 . At present, AI-assisted diagnosis of cataracts, early glaucoma, diabetic retinopathy, age-related macular degeneration, and other ophthalmic diseases is developing rapidly (Table 4) 14-17 . However, studies on AI-assisted diagnosis or screening of dry eye or MGD are relatively rare. Compared with traditional manual diagnosis, the more mature AI diagnosis and segmentation has the following advantages: In terms of image characteristics such as brightness and image contrast, the recognition ability of AI is stronger than that of human eyes; moreover, it does not require the experience of and intervention by imaging physicians. The rate of misdiagnosis caused by subjective factors such as proficiency and fatigue can be avoided. Digital information storage facilitates the development of cases, data exchanges, and cooperation across different regions and fields. Convolutional Neural Network is a deep learning model that has developed rapidly in recent years and received wide attention from various disciplines 18 . It is a pre-feedback neural network structure system that can be used for visual processing. The main structure is composed of multiple groups of units. It is an efficient computing network. In the CNN model, the ResNet model introduces the residual module 19 , which solves the problem of gradient disappearance and model performance degradation as the network depth increases in deep learning. It has become the first choice for image classification tasks at present. IVCM technology has been widely used in the diagnosis of ocular surface diseases. It can be used to observe and evaluate ocular surface inflammation, tissue damage, and nerve distribution at the biological level 20,21 . In 2005, Kobayashi et al. 22 used IVCM for the first time to observe the structural characteristics of the eyelid conjunctiva and lower eyelid meibomian glands in four healthy individuals. Subsequently, some international researchers used IVCM to observe the morphological changes of meibomian glands in healthy subjects and others with MGD, keratitis, Graves ophthalmopathy, and Sjogren’s syndrome, and found that the diameter and density of the acinus, density of inflammatory cells in the acinus, and degree of meibomian gland fibrosis in patients with MGD were significantly abnormal compared with the healthy subjects 23-25 . The degree of keratosis of the meibomian gland openings and the distribution of acinar expansion and atrophy in patients with MGD are helpful indicators to explore the pathogenesis and influencing factors of MGD, and provide a basis for evaluating and detecting the therapeutic effect of MGD from the level of cell morphology. However, there may be some problems in the clinical diagnosis and treatment of MGD by IVCM, such as the large amount of work required for the examination, and low precision of the examination results resulting in a small correlation between the severity of the disease and treatment. These problems restrict the diagnosis and treatment level and the improvement of long-term management effect of DED or MGD. At present, most of the research on AI in MGD mainly focuses on the application of an AI-based diagnosis system to the range of meibomian gland atrophy and to judge whether the patient has MGD according to the characteristics of the gland. Wang et al. 26 used deep learning to automatically segment and calculate the area of meibomian gland atrophy with an accuracy rate of 95.4% and provided quantitative information on the severity of gland atrophy. Koh et al. 27 used slit lamp microscopes equipped with infrared emission filters and infrared cameras combined with machine learning to train and classify based on the length and width characteristics of the glands, and reported a high specificity of 96.1%. In this study, we chose to analyze IVCM images of outpatients from multiple hospitals through an AI-based diagnostic system to assess their consistency, accuracy, sensitivity, and specificity in an attempt to improve the diagnostic rate. Maruoka et al. 28 published a method for detecting MGD by IVCM based on a deep learning model, but the study was mainly a case-control study (MGD and non-MGD). To our knowledge, this study is the first to analyze a six-category MGD network model. We collected a larger number of IVCM images (12,630 images) than previous studies, trained the model with a large number of images, and validated it with data from several hospitals to make the data source more general and the amount of data sufficient. Using Resnet34 migration learning to the MGD dataset, the mean and standard deviation were selected from the parameters obtained by ImageNet dataset. The parameters obtained after extensive pre-training can better improve the accuracy of the model interpretation and the persuasiveness of the study results. The results obtained were as expected, with an accuracy rate of 86.1% or higher for both the internal and external validation sets, which is in high agreement with the ophthalmologist’s diagnosis. The six groups of images had high specificity, ranging from 86.2% (95%CI: 84.9–87.5) to 94.3% (95%CI: 93.5–95.0), suggesting that each group had very low false negatives in the screening and classification diagnosis; this is clinically significant for the classification and diagnosis of MGD. The future integration of AI into confocal microscopy instruments will allow for rapid differentiation of disease types and better detection of disease progression, thus improving the efficiency of ophthalmologists and their ability to manage MGD, and promoting the further development of AI in aiding the diagnosis of ocular surface diseases. However, in this study, our design has some limitations. Since multiple machine learning algorithms were not used (only ResNet34 was used in this study), we cannot guarantee whether the results we obtained are relevant to the chosen algorithm. In addition, although this study covers the data volume of five hospitals, it mainly focuses on the data from South China. In the future, we can collaborate with other hospitals across China to form a complete ' Data Model Diagnosis System'. Once the platform is up and running, doctors' diagnoses and patients' IVCM images will be reused as training sets to continuously update and optimize the diagnostic models. In conclusion, the classification and evaluation of MGD using the DCNN model have good accuracy, which can help clinicians analyze the examination results better and faster, provide a more reliable basis for diagnosis, and support individualized treatment and chronic disease management of MGD and DED. This model is suitable for comprehensive ophthalmology clinics with a large number of visits and a shortage of physicians. It can also be applied to the screening of large populations, follow-up of MGD and DED patients, and observation of efficacy. References 1. Liu Z , et al. Efficacy and Safety of Houttuynia Eye Drops Atomization Treatment for Meibomian Gland Dysfunction-Related Dry Eye Disease: A Randomized, Double-Blinded, Placebo-Controlled Clinical Trial. J Clin Med 9 , (2020). 2. Lemp MA, Crews LA, Bron AJ, Foulks GN, Sullivan BD. Distribution of aqueous-deficient and evaporative dry eye in a clinic-based patient cohort: a retrospective study. Cornea 31 , 472-478 (2012). 3. Kiyat P, Palamar M, Gerceker Turk B, Yagci A. Evaluation of dry eye and Meibomian gland dysfunction in female androgenetic alopecia patients. Int Ophthalmol , (2021). 4. Pedrotti E , et al. In Vivo Confocal Microscopy of the Corneal-Conjunctival Transition in the Evaluation of Epithelial Renewal after SLET. J Clin Med 9 , (2020). 5. Wei S, Ren X, Wang Y, Chou Y, Li X. Therapeutic Effect of Intense Pulsed Light (IPL) Combined with Meibomian Gland Expression (MGX) on Meibomian Gland Dysfunction (MGD). J Ophthalmol 2020 , 3684963 (2020). 6. Ibrahim OM , et al. The efficacy, sensitivity, and specificity of in vivo laser confocal microscopy in the diagnosis of meibomian gland dysfunction. Ophthalmology 117 , 665-672 (2010). 7. Jin C , et al. Development and evaluation of an artificial intelligence system for COVID-19 diagnosis. Nat Commun 11 , 5088 (2020). 8. Ting DSW , et al. Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes. JAMA 318 , 2211-2223 (2017). 9. HONG,X. Expert consensus on diagnosis and treatment of meibomian gland dysfunction in China. Chin J Ophthalmol 53 ,657-661 (2017). 10. Llorella FR, Patow G, Azorin JM. Convolutional neural networks and genetic algorithm for visual imagery classification. Phys Eng Sci Med 43 , 973-983 (2020). 11. Zhang X, Jiang L, Yang D, Yan J, Lu X. Correction to: Urine Sediment Recognition Method Based on Multi-View Deep Residual Learning in Microscopic Image. J Med Syst 44 , 84 (2020). 12. Devalla SK , et al. Glaucoma management in the era of artificial intelligence. Br J Ophthalmol 104 , 301-311 (2020). 13. Moraru AD, Costin D, Moraru RL, Branisteanu DC. Artificial intelligence and deep learning in ophthalmology - present and future (Review). Exp Ther Med 20 , 3469-3473 (2020). 14. Long E , et al. An artificial intelligence platform for the multihospital collaborative management of congenital cataracts. Nature Biomedical Engineering 1 , (2017). 15. Panda R, Puhan NB, Rao A, Mandal B, Padhy D, Panda G. Deep convolutional neural network-based patch classification for retinal nerve fiber layer defect detection in early glaucoma. J Med Imaging (Bellingham) 5 , 044003 (2018). 16. Voets M, Mollersen K, Bongo LA. Reproduction study using public data of: Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. PLoS One 14 , e0217541 (2019). 17. Li Z, He Y, Keel S, Meng W, Chang RT, He M. Efficacy of a Deep Learning System for Detecting Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. Ophthalmology 125 , 1199-1206 (2018). 18. Torsten P, Jan F, Arthur K, Henner S. Using Local Convolutional Neural Networks for Genomic Prediction. Frontiers in Genetics , (2020). 19. Woldegiorgis S, Enqvist A, Baciak J. ResNet and CycleGAN for pulse shape discrimination of He-4 detector pulses: Recovering pulses conventional algorithms fail to label unanimously. Appl Radiat Isot 176 , 109819 (2021). 20. Belmonte C , et al. TFOS DEWS II pain and sensation report. Ocul Surf 15 , 404-437 (2017). 21. Ucar IC, Esen F, Turhan SA, Oguz H, Ulasoglu HC, Aykut V. Corneal neuropathic pain in irritable bowel syndrome: clinical findings and in vivo corneal confocal microscopy. Graefes Arch Clin Exp Ophthalmol , (2021). 22. Kobayashi A, Yoshita T, Sugiyama K. In vivo findings of the bulbar/palpebral conjunctiva and presumed meibomian glands by laser scanning confocal microscopy. Cornea 24 , 985-988 (2005). 23. F F , et al. Facial Seborrheic Keratosis with Unusual Dermoscopic Patterns can be differentiated from other skin malignancies by In Vivo Reflectance Confocal Microscopy. Journal of the European Academy of Dermatology and Venereology : JEADV , (2021). 24. Shengnan C, Yueqi Y, Jin C, Lin Y, Xinghua W, Fagang J. In vivo confocal microscopy assessment of meibomian glands microstructure in patients with Graves' orbitopathy. BMC ophthalmology 21 , (2021). 25. Szalai E, Szucs G, Szamosi S, Aszalos Z, Afra I, Kemeny-Beke A. An in vivo confocal microscopy study of corneal changes in patients with systemic sclerosis. Sci Rep 11 , 11111 (2021). 26. Wang J, Yeh TN, Chakraborty R, Yu SX, Lin MC. A Deep Learning Approach for Meibomian Gland Atrophy Evaluation in Meibography Images. Transl Vis Sci Technol 8 , 37 (2019). 27. Koh YW, Celik T, Lee HK, Petznick A, Tong L. Detection of meibomian glands and classification of meibography images. J Biomed Opt 17 , 086008 (2012). 28. Maruoka S , et al. Deep Neural Network-Based Method for Detecting Obstructive Meibomian Gland Dysfunction With in Vivo Laser Confocal Microscopy. Cornea 39 , 720-725 (2020). 29. Zhongwen L , et al. Preventing corneal blindness caused by keratitis using artificial intelligence. Nature communications 12 , (2021). 30. Zhang K , et al. Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. Nat Biomed Eng 5 , 533-545 (2021). 31. Yang HK, Kim YJ, Sung JY, Kim DH, Kim KG, Hwang J-M. Efficacy for Differentiating Nonglaucomatous Versus Glaucomatous Optic Neuropathy Using Deep Learning Systems. American Journal of Ophthalmology 216 , (2020). 32. Khan SMA, Jens H, Stefan S, E SM, Philipp S. Deep learning-based automatic meibomian gland segmentation and morphology assessment in infrared meibography. Scientific reports 11 , (2021). 33. Daisuke N , et al. Automated detection of a nonperfusion area caused by retinal vein occlusion in optical coherence tomography angiography images using deep learning. PloS one 14 , (2019). Tables Table 1 The total number of six types of MGD, training set ,internal and external validation set cohorts Classification ALL (n=12630) Train set (n=8841) Internal test set (n=3789) External test set (n=12289) Normal group 2896 (22.9%) 1993 (22.5%) 903 (23.8%) 2912 (23.7%) NMGO group 830 (6.6%) 578 (6.5%) 252 (6.7%) 838 (6.8%) MGA group 3585 (28.4%) 2500 (28.3%) 1085 (28.6%) 3493 (28.4%) MGAO group 1745 (13.8%) 1228 (13.9%) 517 (13.6%) 1693 (13.8%) MGO group 3086 (24.4%) 2191 (24.8%) 895 (23.6%) 2866 (23.3%) MGOO group 488 (3.9%) 351 (4.0%) 137 (3.6%) 487 (4.0%) Abbreviations: NMGO: normal with meibomian gland opening; MGA: meibomian gland atrophy; MGAO: meibomian gland atrophy with obstruction; MGO: meibomian gland obstruction; MGOO: meibomian gland obstruction with opening Table 2 Performance metrics of the ensemble DCNN model evaluated on the internal validation set Performance metrics Internal test set (n=3789) Normal group (n=903) NMGO group (n=252) MGA group (n=1085) MGAO group (n=517) MGO group (n=895) MGOO group (n=137) Accuracy (95% CI) 0.893 (0.883 - 0.903) 0.939 (0.931 - 0.947) 0.883 (0.872 - 0.893) 0.914 (0.905 - 0.923) 0.881 (0.871 - 0.892) 0.933 (0.925 - 0.941) Sensitivity (95% CI) 0.893 (0.871 - 0.912) 0.889 (0.843 - 0.925) 0.934 (0.917 - 0.948) 0.919 (0.892 - 0.941) 0.874 (0.850 - 0.895) 0.876 (0.809 - 0.926) Specificity (95% CI) 0.893 (0.881 - 0.904) 0.943 (0.935 - 0.950) 0.862 (0.849 - 0.875) 0.914 (0.903 - 0.923) 0.884 (0.872 - 0.895) 0.936 (0.927 - 0.943) Precision (95% CI) 0.723 (0.696 - 0.749) 0.526 (0.477 - 0.574) 0.731 (0.707 - 0.755) 0.627 (0.591 - 0.661) 0.699 (0.672 - 0.726) 0.338 (0.289 - 0.390) Negative predictive value (95% CI) 0.964 (0.956 - 0.970) 0.992 (0.988 - 0.994) 0.970 (0.962 - 0.976) 0.986 (0.981 - 0.990) 0.958 (0.949 - 0.965) 0.995 (0.992 - 0.997) Kappa § 0.727 0.630 0.735 0.696 0.698 0.460 F 1 † 0.799 0.661 0.820 0.745 0.777 0.488 AUROC (95% CI) 0.957 (0.951 - 0.964) 0.965 (0.954 - 0.976) 0.959 (0.954 - 0.965) 0.971 (0.965 - 0.977) 0.947 (0.939 - 0.955) 0.963 (0.947 - 0.978) Abbreviations: NMGO: normal with meibomian gland opening; MGA: meibomian gland atrophy; MGAO: meibomian gland atrophy with obstruction; MGO: meibomian gland obstruction; MGOO: meibomian gland obstruction with opening;AUROC: Area Under the Receiver Operating Characteristic curve Table 3 Performance metrics of the ensemble DCNN model evaluated on the external validation set Performance metrics External test set (n=12289) Normal group (n=2912) NMGO group (n=838) MGA group (n=3493) MGAO group (n=1693) MGO group (n=2866) MGOO group (n=487) Accuracy (95% CI) 0.894 (0.889 - 0.900) 0.929 (0.925 - 0.934) 0.885 (0.879 - 0.890) 0.916 (0.911 - 0.921) 0.881 (0.875 - 0.886) 0.939 (0.934 - 0.943) Sensitivity (95% CI) 0.901 (0.890 - 0.912) 0.906 (0.884 - 0.925) 0.940 (0.931 - 0.948) 0.926 (0.912 - 0.938) 0.893 (0.881 - 0.904) 0.825 (0.789 - 0.858) Specificity (95% CI) 0.892 (0.886 - 0.898) 0.931 (0.926 - 0.936) 0.863 (0.856 - 0.870) 0.914 (0.909 - 0.920) 0.877 (0.870 - 0.884) 0.943 (0.939 - 0.947) Precision (95% CI) 0.722 (0.707 - 0.736) 0.490 (0.465 - 0.515) 0.732 (0.718 - 0.744) 0.633 (0.614 - 0.652) 0.689 (0.673 - 0.703) 0.375 (0.346 - 0.405) Negative predictive value (95% CI) 0.967 (0.963 - 0.970) 0.993 (0.991 - 0.994) 0.973 (0.969 - 0.977) 0.987 (0.985 - 0.989) 0.964 (0.960 - 0.968) 0.992 (0.991 - 0.994) Kappa § 0.731 0.601 0.739 0.703 0.698 0.488 F 1 † 0.802 0.636 0.823 0.752 0.777 0.516 AUROC (95% CI) 0.960 (0.957 - 0.963) 0.971 (0.966 - 0.976) 0.961 (0.958 - 0.964) 0.974 (0.971 - 0.977) 0.954 (0.950 - 0.958) 0.951 (0.941 - 0.961) Abbreviations: NMGO: normal with meibomian gland opening; MGA: meibomian gland atrophy; MGAO: meibomian gland atrophy with obstruction; MGO: meibomian gland obstruction; MGOO: meibomian gland obstruction with opening; AUROC: Area Under the Receiver Operating Characteristic curve Table 4 Application of deep learning in ophthalmic diseases Author Year Diseases Deep learning algorithms Refs. Li et al 2021 Keratitis DenseNet121, Inception-v3, ResNet-50 29 Zhang et al 2021 Chronic kidney disease and diabetic fundus disease ResNet-50 30 Yang et al 2020 Glaucomatous Optic Neuropathy ResNet-50 31 Khan et al 2021 Meibomian gland dysfunction ResNet-v2 32 Daisuke et al 2019 Branch Retinal Vein Occlusion Visual Geometry Group (VGG)-16 33 Declarations Acknowledgements Not applicable. Foundation item ： National Natural Science Foundation（No: 82160195）; Central Government Guides Local Science and Technology Development Foundation(No: 20211ZDG02003); Key Research Foundation of Jiangxi Province (No: 20181BBG70004, 20203BBG73059); Excellent Talents Development Project of Jiangxi Province（No: 20192BCBL23020）; Natural Science Foundation of Jiangxi Province（No: 20181BAB205034）; Grassroots Health Appropriate Technology “Spark Promotion Plan” Project of Jiangxi Province(No:20188003); Health Development Planning Commission Science Foundation of Jiangxi Province (No: 20201032,202130210) Availability of data and materials The datasets used and/or analyzed during the present study are available from the corresponding author on reasonable request. Ethical approval and consent to participate The study methods and protocols were approved by the Medical Ethics Committee of the First Affiliated Hospital of Nanchang University (Nanchang, China) and followed the principles of the Declaration of Helsinki. All subjects were notified of the objectives and content of the study and latent risks, and then provided written informed consent to participate. Patient consent for publication Not applicable. Competing interests This study did not receive any industrial support. The authors have no competing interests to declare regarding this study Author contributions All the authors contributed to this manuscript. Yi Shao and Yi-Chen Yang are responsible for conceiving and designing the work, acquiring data and writing the manuscript; Hui Zhao and Wen-Qing Shi played an important role in interpreting the results; Xu-Lin Liao, Ting Su and Rong-Bin Liang Min helped in acquiring data and giving some advice; Qiu-Yu Li and Qian-Min Ge contributed to the application to the ethics committee; Hui-Ye Su and Yi-Cong Pan revised the manuscript, and Xiang-Chun Li gave valuable guidance at every stage and approved the final version. Additional Declarations There is NO Competing Interest. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-936418","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":54486489,"identity":"f145ff05-455c-42a1-8f7e-1c8246a9f131","order_by":0,"name":"Yi Shao","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Yi","middleName":"","lastName":"Shao","suffix":""},{"id":54486490,"identity":"40d0f381-d3a0-4a19-8691-0102f291d589","order_by":1,"name":"Yichen Yang","email":"","orcid":"","institution":"Tianjin Medical University Cancer Institute and Hospital","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Yichen","middleName":"","lastName":"Yang","suffix":""},{"id":54486491,"identity":"ec0abb16-19c9-4e6d-b49e-e5df413652fd","order_by":2,"name":"Hui Zhao","email":"","orcid":"","institution":"Shanghai Jiao Tong University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Zhao","suffix":""},{"id":54486492,"identity":"2e202adf-8a86-4c2d-ae7c-e458f371b473","order_by":3,"name":"Wen-Qing Shi","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Wen-Qing","middleName":"","lastName":"Shi","suffix":""},{"id":54486493,"identity":"227c5b75-0044-4c5f-a8c7-849a0b605365","order_by":4,"name":"Xu-Lin Liao","email":"","orcid":"https://orcid.org/0000-0002-2800-9838","institution":"The Chinese University of Hong Kong","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xu-Lin","middleName":"","lastName":"Liao","suffix":""},{"id":54486494,"identity":"b2127f4c-85c0-49d0-a233-08373b81854c","order_by":5,"name":"Ting Su","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Ting","middleName":"","lastName":"Su","suffix":""},{"id":54486495,"identity":"06ed743c-7c0c-43dd-bb18-21c3f7ca0c4b","order_by":6,"name":"Rong-Bin Liang","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Rong-Bin","middleName":"","lastName":"Liang","suffix":""},{"id":54486496,"identity":"6156da8c-9f40-4b00-b291-41829790136f","order_by":7,"name":"Qiu-Yu Li","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Qiu-Yu","middleName":"","lastName":"Li","suffix":""},{"id":54486497,"identity":"54810e6a-6a2f-4f61-bddc-ec5b480c61bc","order_by":8,"name":"Qian-Min Ge","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Qian-Min","middleName":"","lastName":"Ge","suffix":""},{"id":54486498,"identity":"bcc1a4a1-d131-4a34-8207-551c5538e19c","order_by":9,"name":"Hui-Ye Shu","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Hui-Ye","middleName":"","lastName":"Shu","suffix":""},{"id":54486499,"identity":"72a0956e-0652-4362-b27e-60582686036c","order_by":10,"name":"Yi-Cong Pan","email":"","orcid":"","institution":"Nanchang University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Yi-Cong","middleName":"","lastName":"Pan","suffix":""},{"id":54486500,"identity":"42dd50da-c2a3-4677-ae08-161ae56429c8","order_by":11,"name":"Xiangchun Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5UlEQVRIiWNgGAWjYBACPmYGhgMghgEDA+MDiFgCfi1sSFqYDQ4QpQXGAGphkyBOCzuP4YGfO2rlzSVyj1V/zDnMwM+eY8Dwcwc+h7ElHOw9c9xw54y8tBsHtx1mkOx5Y8DYewafFuYDB3jbjiUY3MgxA2sBMgyYGdvwaWFsOPgXqqUApMWesBbmA4d522rAWhjAtkgQ1MKWcFi27YDhhjNvjCXObkvnkTjzrOBgLx4t/PxnjD++bauTNzieY/ihcpu1HH978sYHP/FogYLDcBYPiDhAUAMDQx0RakbBKBgFo2DEAgAJb1Ju/dCh9gAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-2230-2152","institution":"Tianjin Medical University Cancer Institute and Hospital","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Xiangchun","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2021-09-25 04:10:48","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-936418/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-936418/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":14879303,"identity":"4063aee6-7c0d-422a-b99b-25ac2899e988","added_by":"auto","created_at":"2021-10-25 17:17:54","extension":"tif","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":6333784,"visible":true,"origin":"","legend":"Representative MG images obtained under in vivo confocal microscope. \nNotes: A:Normal group;B:Normal with meibomian gland opening group .C:Meibomian gland atrophy group.D:Meibomian gland obstruction group.E:Meibomian gland atrophy with obstruction group .F:Meibomian gland obstruction with opening group.Blue arrow: Normal acinar; Green arrow: Meibomian gland opening; Red arrow: Atrophy of acinar wall; Yellow arrow: Acinar containing blocked lipid.\n","description":"","filename":"fig1.tif","url":"https://assets-eu.researchsquare.com/files/rs-936418/v1/2ddef2a9cbfebd9b3e9b0f1f.tif"},{"id":14879305,"identity":"eabefd0e-939f-4b2d-81f4-1ae9c0893728","added_by":"auto","created_at":"2021-10-25 17:17:54","extension":"tif","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3567135,"visible":true,"origin":"","legend":"Flow chart of deep learning model development and internal validation process for automatic diagnosis of meibomian gland dysfunction.\nNMGO: Normal with meibomian gland opening; MGA: Meibomian gland atrophy; MGO: Meibomian gland obstruction; MGAO: Meibomian gland atrophy with obstruction; MGOO: Meibomian gland obstruction with opening; MGD: Meibomian gland dysfunction; AUROC: Area under the receiver operating characteristic curve.\n","description":"","filename":"fig2.tif","url":"https://assets-eu.researchsquare.com/files/rs-936418/v1/508dff4f81034cf5b8fd003d.tif"},{"id":14879304,"identity":"de31398a-d0fc-489c-852b-bb5fe66c5929","added_by":"auto","created_at":"2021-10-25 17:17:54","extension":"tif","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1117019,"visible":true,"origin":"","legend":"The ROC curve of the six classification internal validation sets.\nNMGO: Normal with meibomian gland opening; MGA: Meibomian gland atrophy; MGO: Meibomian gland obstruction; MGAO: Meibomian gland atrophy with obstruction; MGOO: Meibomian gland obstruction with opening; AUROC: Area under the receiver operating characteristic curve.\n","description":"","filename":"fig3.tif","url":"https://assets-eu.researchsquare.com/files/rs-936418/v1/cedcd2f733f1819ea91d1a32.tif"},{"id":14879302,"identity":"6f2515b4-5d70-49d3-97c4-f8fe1462f15e","added_by":"auto","created_at":"2021-10-25 17:17:54","extension":"tif","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1027517,"visible":true,"origin":"","legend":"The ROC curve of the six classification external validation sets.\nNMGO: Normal with meibomian gland opening; MGA: Meibomian gland atrophy; MGO: Meibomian gland obstruction; MGAO: Meibomian gland atrophy with obstruction; MGOO: Meibomian gland obstruction with opening; AUROC: Area under the receiver operating characteristic curve.\n","description":"","filename":"Figure4.tif","url":"https://assets-eu.researchsquare.com/files/rs-936418/v1/7c945274ab044860c6cf42b7.tif"},{"id":51310542,"identity":"59e2ad60-514d-413b-ab91-50ea19524c6a","added_by":"auto","created_at":"2024-02-19 10:53:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3970856,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-936418/v1/d9e8748c-74bc-4d7a-9532-b8ae7ef17f31.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Detection of Meibomian Gland Dysfunction by in vivo Confocal Microscopy Based on Deep Convolutional Neural Network","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMeibomian glands are the largest sebaceous glands in the human body. They are distributed in the upper and lower meibomian layers, with 30\u0026ndash;40 on the upper eyelid and 20\u0026ndash;30 on the lower eyelid and are perpendicular to the eyelid margin. The meibomian gland can synthesize, store, and secrete lipids, and these secreted lipids constitute the lipid layer of the tear film. Meibomian gland dysfunction (MGD) is a chronic, non-specific inflammation characterized by meibomian gland duct obstruction or abnormal meibomian gland secretions. In MGD, the glandular lipid secretion is impaired, the ocular surface cannot maintain the stability of the tear film, the tear film evaporates rapidly, and the tear osmotic pressure increases, which can lead to dry eye disease (DED)\u003csup\u003e1,2\u003c/sup\u003e. With increasing age, the acinar epithelial cells of the meibomian gland shrink, leading to large-scale atrophy of the meibomian glands. As these glands are non-renewable, MGD reduces lipid secretion\u003csup\u003e3\u003c/sup\u003e. With an increase in the aging population, environmental pollution, and widespread use of video terminals, the incidence of MGD has gradually increased and become a disease of global proportions. The clinical symptoms of MGD are similar to those of DED and include dry eyes, eye irritation, blurred vision, and increased secretions. As the symptoms of MGD are not specific, it is easy to be missed, misdiagnosed, and mistreated. Therefore, objective physical signs can provide a reference for diagnosis, including the loss of meibomian glands, abnormal secretion of meibomian glands, and changes in the morphology of the eyelid margin.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;In vivo confocal microscopy (IVCM) is a new type of high-resolution, non-invasive ocular surface imaging detector. The laser focusing principle of IVCM is used to scan each point of the detected layer in the confocal plane. In IVCM, the formation of ocular surface tissues can be observed at multiple levels, and the cell level of tissue morphology and changes can be photographed in real time\u003csup\u003e4\u003c/sup\u003e. Hence, IVCM has real-time, non-invasive, three-dimensional advantages. In recent years, as MGD has received increasing attention, there have been an increasing number of researches on the application of IVCM in the meibomian glands\u003csup\u003e5\u003c/sup\u003e. IVCM can facilitate clear visualization of the characteristics of the meibomian glands, such as acinus, acinar stroma, gland secretions, periadenositis cells, and meibomian gland openings and classify them for standardized treatment. Ibrahim et al.\u003csup\u003e6\u003c/sup\u003e obtained IVCM images and analyzed the density of meibomian acinar units and the maximum and minimum diameters of the meibomian acinars; they concluded that IVCM has a high level of specificity and sensitivity in the diagnosis of MGD. However, because of the limitations of clinical manpower and material resources, the diagnosis of MGD-related eye diseases still has limitations at present, such as the limited surgical skills of junior doctors and the huge workload of manual classification, which are not conducive to the development of individualized disease treatment and chronic disease management. Recent advancements in the field of artificial intelligence (AI) have led to its increased application in disease assessment and diagnosis; further, AI has also become a research focus for clinical professionals. Analyzing MGD IVCM images data with high-performance AI algorithm can help distinguish different types of patients, greatly increase the work efficiency of clinicians, and save time by formulating different treatment plans for patients, in order to reduce the social and medical burden and aid with the new trend of personalized and refined MGD diagnosis and treatment. In recent years, the combined deep convolutional neural network (DCNN) model has been applied to the diagnosis of various medical diseases\u003csup\u003e7,8\u003c/sup\u003e. We propose a method that combines the DCNN algorithm and IVCM diagnosis for early diagnosis and classification of MGD.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy Subjects\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Institutional Review Board of the Ethics Committee of the First Affiliated Hospital of Nanchang University approved this retrospective study, and all methods complied with the tenets of the Helsinki Declaration.\u003c/p\u003e\n\u003cp\u003eThis was a single-center, clinical study. From January 2018 to June 2021, we extracted 12,630 IVCM images from the IVCM database of three hospitals, namely the First Affiliated Hospital of Nanchang University, Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine, and Shanghai General Hospital. Three ophthalmologists checked the quality of the images and performed clinical diagnosis and classification based on the 2017 Chinese Expert Consensus on the Diagnosis and Treatment of MGD\u003csup\u003e9\u003c/sup\u003e. The extracted data included the following: normal group (n=2896), normal with meibomian gland opening (NMGO) group (n=830), meibomian gland atrophy (MGA) group (n=3585), meibomian gland atrophy with obstruction (MGAO) group (n=1745), meibomian gland obstruction (MGO) group (n=3086), and meibomian gland obstruction with opening (MGOO) group (n=488). The six types of IVCM classification pictures can be seen in Figure 1. According to random grouping, 70% (n=8,841) and 30% (n=3,789) images were allocated to the training and internal validation sets, respectively. Simultaneously, we used 12,289 images from Nanchang University Affiliated Eye Hospital and Shanghai AIER Eye Hospital as an external verification set to verify the performance of the model. The DCNN classification model that maps the input features (such as image pixels) to the corresponding output labels (MGD of different classifications) was used to train the deep learning algorithm. We used the ResNet34 network model for classification and initial training(120 rounds of training) and developed and trained the DCNN model on the training set. We enhanced the images from the training data set using brightness, gamma correction, histogram equalization, noise addition, and inversion. The process flow chart of this research is shown in Figure 2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClassification Standard\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNormal group: meibomian glands and ducts are normal.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNMGO group: full meibomian glands with opening.\u003c/p\u003e\n\u003cp\u003eMGA group: tire-like meibomian gland epithelial cells are absent, the acinar wall disappears, becomes smaller, or the fiber cord changes.\u003c/p\u003e\n\u003cp\u003eMGO group: enlarged acinus and blocked meibum in the acinus.\u003c/p\u003e\n\u003cp\u003eMGAO group: shows the characteristics of both the atrophy group and obstruction group.\u003c/p\u003e\n\u003cp\u003eMGOO group: normal meibomian gland opening with highly dilated meibomian gland secretory tube and blocked meibomian gland acinus.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIn Vivo Confocal Microscope Image Acquisition\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe IVCM (HRT-3, RCM module, Hidelberg Engineering, Germany) parameters were as follows: laser wavelength, 670 nm; observation field, 400 um*400 um; resolution, 384 pixels * 384 pixels; magnification, 800 times; and axis resolution, is 1 um. Use 0.5% proparacaine eye drops (American Alcon company) for topical anesthesia,and apply Carbomer Eye Gel (Shandong Bausch \u0026amp; Lomb Freda Pharmaceutical Co., Ltd.) on the surface of the IVCM contact lens. The disposable corneal contact cover is attached to the contact lens. To make the image clearer, the outer layer of the contact sleeve is evenly smeared with Carbomer Eye Gel. The examinee\u0026rsquo;s chin and forehead are fixed at the corresponding positions of the IVCM, and the eyelid of the examinee\u0026rsquo;s eye is turned over. The examiner fixes the eyelid with one hand to keep it in a state of eversion, and pushes the IVCM lens forward with the other hand to make it touch the edge of the examinee\u0026rsquo;s eyelid, and adjusts the scan depth, starting from the eyelid margin and gradually scanning the meibomian glands in the fornix, from the nasal side of the lesion to the temporal side, while taking IVCM images.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Building\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConvolutional neural network (CNN) is a multi-layer neural network with fault tolerance and is easy to train and optimize\u003csup\u003e10\u003c/sup\u003e. The residual network (ResNet) is a type of CNN, the network is connected by adding one operation channel in every two layers to form a residual block to reach a jump in the whole network, avoiding the gradient descent problem and network degradation problem caused by the multiplication of very small parameters in each layer\u003csup\u003e11\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn this study, we chose a RESNET network (ResNet34) composed of 34 convolution layers to classify the MGD images. Image features were extracted using sixty-four 7\u0026times;7 convolution kernels, and then a 3\u0026times;3 maximum down-sampling operation with a stride of 2 was performed. Then, input the result into four residual groups with three, four, six, and three residual units, where each residual unit has two 3\u0026times;3 convolutional layers with a stride of 2. After each convolutional layer, the linear rectification function (ReLU) is used as the activation function. It is mapped to the probability of each category by the fully connected layer and the Softmax activation function.\u003c/p\u003e\n\u003cp\u003eIn this experiment, a desktop computer was used as the experimental environment: the Central Processing Unit (CPU) is Ryzen5 1500X (Advanced Micro Devices, Inc.), the memory is 16 G, the graphic processing unit (GPU) is Nvidia GeForce GTX1060, and the graphics card memory is 6 G. We used the PyTorch (https://py-torch.org) Python framework based on the Windows 10 system to develop and train the MGD image recognition model. First, we used the pre-trained generation weight initialization network on the ImageNet (http://pytorch.org/docs/stable/torchvision/models.html) dataset, and then transferred the model network to MGD. On the data set, the modeling process is divided into three parts: image preprocessing, image feature extraction and classification of MGD IVCM, and active learning. The optimizer is Adam. The model training parameters are as follows: batch size=16, epoch=120, and learning rate=0.1; the other parameters are default.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Evaluation and Statistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter the MGD IVCM image recognition model is trained, the performance of the trained model needs to be evaluated, and the results of classification by three ophthalmologists from the First Affiliated Hospital of Nanchang University were used as the gold standard to calculate the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, negative predictive value (NPV), Kappa coefficient, and F1 scores. Four statistical parameters are involved: true positive (TP), false positive (FP), true negative (TN), and false negative (FN). The calculation formula is as follows: accuracy=(TP+TN)/(TP+FP+FN+TN), specificity=TN/N (negative), precision (P)=TP/(TP+FP), sensitivity=TP/P, and NPV=TN/(TN+FN). The ROC curve was plotted with specificity as the abscissa and sensitivity as the ordinate; the larger the AUROC value, the better the classification performance of the model. The Kappa test was used to compare the consistency of the results of the artificial intelligence diagnosis group and the expert group and to evaluate the accuracy of the multi-classification model. All the above indicators were obtained using the R software (version 6.078).\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eData Set Characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 24,919 images from five hospitals were used to train and verify the performance of the DCNN. From the First Affiliated Hospital of Nanchang University, Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine, and Shanghai General Hospital, the training set consisted of 8841 images (1993 images in the normal group, 578 images in the NMGO group, 2500 images in the MGA group, 1228 images in the MGAO group, 2191 images in the MGO group, and 351 images in the MGOO group). Additionally, 3789 internal validation sets (903 in the normal group, 252 in the NMGO group, 1085 in the MGA group, 517 in the MGAO group, 895 in the MGO group, and 137 in the MGOO group), and a total of 12289 images were collected from the external validation set of Nanchang University Affiliated Eye Hospital and Shanghai AIER Eye Hospital(2912 in the normal group, 838 in the NMGO group, 3493 in the MGA group, 1693 in the MGAO group, 2866 in the MGO group, and 487 in the MGOO group). The total number, training set, internal and external validation set cohorts of the six types of MGD are summarized in Table 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance of Deep Learning Algorithms\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe used the DCNN algorithm ResNet34 to train the classification model on the same training data set. Based on the 30% of randomly selected pictures for internal verification, the accuracy, sensitivity, specificity, precision, and AUROC parameters of the classification were calculated. The accuracy of the internal validation set test of the six groups of pictures ranged from 88.1% (95%CI: 87.1\u0026ndash;89.2) to 93.9% (95%CI: 93.1\u0026ndash;94.7); sensitivity, from 87.4% (95%CI: 85.0\u0026ndash;89.5) to 93.4% (95%CI: 85.0\u0026ndash;89.5); specificity, from 86.2% (95%CI: 84.9\u0026ndash;87.5) to 94.3% (95%CI: 93.5\u0026ndash;95.0); precision, from 33.8% (95%CI: 28.9\u0026ndash;39.0) to 73.1% (95%CI: 70.7\u0026ndash;75.5); AUROC, from 0.947 (95%CI: 0.939\u0026ndash;0.955) to 0.971 (95%CI: 0.965\u0026ndash;0.977). The accuracy of the six-category external validation set to verify the classification model ranged from 88.1% (95%CI: 87.5\u0026ndash;88.6) to 93.9% (95%CI: 93.4\u0026ndash;94.3); sensitivity, from 82.5% (95%CI: 78.9\u0026ndash;85.8) to 94.0% (95%CI: 93.1\u0026ndash;94.8); specificity, from 86.3% (95%CI: 85.6\u0026ndash;87.0) to 94.3% (95%CI: 93.9\u0026ndash;94.7); and AUROC, from 0.951 (95%CI: 0.941\u0026ndash;0.961) to 0.971 (95%CI: 0.966\u0026ndash;0.976). The built DCNN model internal and external validation set had high negative predictive value, Kappa coefficient, and F1 score value. The internal validation set test results are shown in the following table (Table 2). The test results of the external validation set are shown in Table 3. The ROC curve of the six classification validation sets is shown in Figures 3 and 4.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we aimed to evaluate the performance of convolutional neural networks for detecting MGD using a large number of confocal microscopy images with different classifications. The results show that the combination of the DCNN model and confocal microscopy images that we tested in the validation set can finely classify MGD patients in a realistic setting with high accuracy, specificity, and AUROC values, compared to manual classification by ophthalmologists.\u003c/p\u003e\n\u003cp\u003eIn recent years, with the rapid development of AI, its application in clinical practice has become increasingly widespread\u003csup\u003e12,13\u003c/sup\u003e. At present, AI-assisted diagnosis of cataracts, early glaucoma, diabetic retinopathy, age-related macular degeneration, and other ophthalmic diseases is developing rapidly (Table 4)\u003csup\u003e14-17\u003c/sup\u003e. However, studies on AI-assisted diagnosis or screening of dry eye or MGD are relatively rare. Compared with traditional manual diagnosis, the more mature AI diagnosis and segmentation has the following advantages: In terms of image characteristics such as brightness and image contrast, the recognition ability of AI is stronger than that of human eyes; moreover, it does not require the experience of and intervention by imaging physicians. The rate of misdiagnosis caused by subjective factors such as proficiency and fatigue can be avoided. Digital information storage facilitates the development of cases, data exchanges, and cooperation across different regions and fields.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConvolutional Neural Network is a deep learning model that has developed rapidly in recent years and received wide attention from various disciplines\u003csup\u003e18\u003c/sup\u003e. It is a pre-feedback neural network structure system that can be used for visual processing. The main structure is composed of multiple groups of units. It is an efficient computing network. In the CNN model, the ResNet model introduces the residual module\u003csup\u003e19\u003c/sup\u003e, which solves the problem of gradient disappearance and model performance degradation as the network depth increases in deep learning. It has become the first choice for image classification tasks at present.\u003c/p\u003e\n\u003cp\u003eIVCM technology has been widely used in the diagnosis of ocular surface diseases. It can be used to observe and evaluate ocular surface inflammation, tissue damage, and nerve distribution at the biological level\u003csup\u003e20,21\u003c/sup\u003e. In 2005, Kobayashi et al.\u003csup\u003e22\u003c/sup\u003e used IVCM for the first time to observe the structural characteristics of the eyelid conjunctiva and lower eyelid meibomian glands in four healthy individuals. Subsequently, some international researchers used IVCM to observe the morphological changes of meibomian glands in healthy subjects and others with MGD, keratitis, Graves ophthalmopathy, and Sjogren\u0026rsquo;s syndrome, and found that the diameter and density of the acinus, density of inflammatory cells in the acinus, and degree of meibomian gland fibrosis in patients with MGD were significantly abnormal compared with the healthy subjects\u003csup\u003e23-25\u003c/sup\u003e. The degree of keratosis of the meibomian gland openings and the distribution of acinar expansion and atrophy in patients with MGD are helpful indicators to explore the pathogenesis and influencing factors of MGD, and provide a basis for evaluating and detecting the therapeutic effect of MGD from the level of cell morphology. However, there may be some problems in the clinical diagnosis and treatment of MGD by IVCM, such as the large amount of work required for the examination, and low precision of the examination results resulting in a small correlation between the severity of the disease and treatment. These problems restrict the diagnosis and treatment level and the improvement of long-term management effect of DED or MGD.\u003c/p\u003e\n\u003cp\u003eAt present, most of the research on AI in MGD mainly focuses on the application of an AI-based diagnosis system to the range of meibomian gland atrophy and to judge whether the patient has MGD according to the characteristics of the gland. Wang et al.\u003csup\u003e26\u003c/sup\u003e used deep learning to automatically segment and calculate the area of meibomian gland atrophy with an accuracy rate of 95.4% and provided quantitative information on the severity of gland atrophy. Koh et al.\u003csup\u003e27\u003c/sup\u003e used slit lamp microscopes equipped with infrared emission filters and infrared cameras combined with machine learning to train and classify based on the length and width characteristics of the glands, and reported a high specificity of 96.1%. In this study, we chose to analyze IVCM images of outpatients from multiple hospitals through an AI-based diagnostic system to assess their consistency, accuracy, sensitivity, and specificity in an attempt to improve the diagnostic rate. Maruoka et al.\u003csup\u003e28\u003c/sup\u003e published a method for detecting MGD by IVCM based on a deep learning model, but the study was mainly a case-control study (MGD and non-MGD). To our knowledge, this study is the first to analyze a six-category MGD network model. We collected a larger number of IVCM images (12,630 images) than previous studies, trained the model with a large number of images, and validated it with data from several hospitals to make the data source more general and the amount of data sufficient. Using Resnet34 migration learning to the MGD dataset, the mean and standard deviation were selected from the parameters obtained by ImageNet dataset. The parameters obtained after extensive pre-training can better improve the accuracy of the model interpretation and the persuasiveness of the study results. The results obtained were as expected, with an accuracy rate of 86.1% or higher for both the internal and external validation sets, which is in high agreement with the ophthalmologist\u0026rsquo;s diagnosis. The six groups of images had high specificity, ranging from 86.2% (95%CI: 84.9\u0026ndash;87.5) to 94.3% (95%CI: 93.5\u0026ndash;95.0), suggesting that each group had very low false negatives in the screening and classification diagnosis; this is clinically significant for the classification and diagnosis of MGD. The future integration of AI into confocal microscopy instruments will allow for rapid differentiation of disease types and better detection of disease progression, thus improving the efficiency of ophthalmologists and their ability to manage MGD, and promoting the further development of AI in aiding the diagnosis of ocular surface diseases.\u003c/p\u003e\n\u003cp\u003eHowever, in this study, our design has some limitations. Since multiple machine learning algorithms were not used (only ResNet34 was used in this study), we cannot guarantee whether the results we obtained are relevant to the chosen algorithm. In addition, although this study covers the data volume of five hospitals, it mainly focuses on the data from South China. In the future, we can collaborate with other hospitals across China to form a complete \u0026apos; Data Model Diagnosis System\u0026apos;. Once the platform is up and running, doctors\u0026apos; diagnoses and patients\u0026apos; IVCM images will be reused as training sets to continuously update and optimize the diagnostic models.\u003c/p\u003e\n\u003cp\u003eIn conclusion, the classification and evaluation of MGD using the DCNN model have good accuracy, which can help clinicians analyze the examination results better and faster, provide a more reliable basis for diagnosis, and support individualized treatment and chronic disease management of MGD and DED. This model is suitable for comprehensive ophthalmology clinics with a large number of visits and a shortage of physicians. It can also be applied to the screening of large populations, follow-up of MGD and DED patients, and observation of efficacy.\u003c/p\u003e"},{"header":"References","content":"\u003cp\u003e1.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Liu Z\u003cem\u003e, et al.\u003c/em\u003e Efficacy and Safety of Houttuynia Eye Drops Atomization Treatment for Meibomian Gland Dysfunction-Related Dry Eye Disease: A Randomized, Double-Blinded, Placebo-Controlled Clinical Trial. \u003cem\u003eJ Clin Med\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, \u0026nbsp;(2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Lemp MA, Crews LA, Bron AJ, Foulks GN, Sullivan BD. Distribution of aqueous-deficient and evaporative dry eye in a clinic-based patient cohort: a retrospective study. \u003cem\u003eCornea\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 472-478 (2012).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Kiyat P, Palamar M, Gerceker Turk B, Yagci A. Evaluation of dry eye and Meibomian gland dysfunction in female androgenetic alopecia patients. \u003cem\u003eInt Ophthalmol\u003c/em\u003e, \u0026nbsp;(2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e4.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Pedrotti E\u003cem\u003e, et al.\u003c/em\u003e In Vivo Confocal Microscopy of the Corneal-Conjunctival Transition in the Evaluation of Epithelial Renewal after SLET. \u003cem\u003eJ Clin Med\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, \u0026nbsp;(2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e5.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Wei S, Ren X, Wang Y, Chou Y, Li X. Therapeutic Effect of Intense Pulsed Light (IPL) Combined with Meibomian Gland Expression (MGX) on Meibomian Gland Dysfunction (MGD). \u003cem\u003eJ Ophthalmol\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, 3684963 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e6.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Ibrahim OM\u003cem\u003e, et al.\u003c/em\u003e The efficacy, sensitivity, and specificity of in vivo laser confocal microscopy in the diagnosis of meibomian gland dysfunction. \u003cem\u003eOphthalmology\u003c/em\u003e \u003cstrong\u003e117\u003c/strong\u003e, 665-672 (2010).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e7.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Jin C\u003cem\u003e, et al.\u003c/em\u003e Development and evaluation of an artificial intelligence system for COVID-19 diagnosis. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 5088 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e8.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Ting DSW\u003cem\u003e, et al.\u003c/em\u003e Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes. \u003cem\u003eJAMA\u003c/em\u003e \u003cstrong\u003e318\u003c/strong\u003e, 2211-2223 (2017).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e9.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;HONG,X. Expert consensus on diagnosis and treatment of meibomian gland dysfunction in China. \u003cem\u003eChin J Ophthalmol\u003c/em\u003e \u003cstrong\u003e53\u003c/strong\u003e,657-661 (2017).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e10.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Llorella FR, Patow G, Azorin JM. Convolutional neural networks and genetic algorithm for visual imagery classification. \u003cem\u003ePhys Eng Sci Med\u003c/em\u003e \u003cstrong\u003e43\u003c/strong\u003e, 973-983 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e11.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Zhang X, Jiang L, Yang D, Yan J, Lu X. Correction to: Urine Sediment Recognition Method Based on Multi-View Deep Residual Learning in Microscopic Image. \u003cem\u003eJ Med Syst\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, 84 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e12.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Devalla SK\u003cem\u003e, et al.\u003c/em\u003e Glaucoma management in the era of artificial intelligence. \u003cem\u003eBr J Ophthalmol\u003c/em\u003e \u003cstrong\u003e104\u003c/strong\u003e, 301-311 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e13.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Moraru AD, Costin D, Moraru RL, Branisteanu DC. Artificial intelligence and deep learning in ophthalmology - present and future (Review). \u003cem\u003eExp Ther Med\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, 3469-3473 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e14.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Long E\u003cem\u003e, et al.\u003c/em\u003e An artificial intelligence platform for the multihospital collaborative management of congenital cataracts. \u003cem\u003eNature Biomedical Engineering\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, \u0026nbsp;(2017).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e15.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Panda R, Puhan NB, Rao A, Mandal B, Padhy D, Panda G. Deep convolutional neural network-based patch classification for retinal nerve fiber layer defect detection in early glaucoma. \u003cem\u003eJ Med Imaging (Bellingham)\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 044003 (2018).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e16.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Voets M, Mollersen K, Bongo LA. Reproduction study using public data of: Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, e0217541 (2019).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e17.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Li Z, He Y, Keel S, Meng W, Chang RT, He M. Efficacy of a Deep Learning System for Detecting Glaucomatous Optic Neuropathy Based on Color Fundus Photographs. \u003cem\u003eOphthalmology\u003c/em\u003e \u003cstrong\u003e125\u003c/strong\u003e, 1199-1206 (2018).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e18.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Torsten P, Jan F, Arthur K, Henner S. Using Local Convolutional Neural Networks for Genomic Prediction. \u003cem\u003eFrontiers in Genetics\u003c/em\u003e, \u0026nbsp;(2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e19.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Woldegiorgis S, Enqvist A, Baciak J. ResNet and CycleGAN for pulse shape discrimination of He-4 detector pulses: Recovering pulses conventional algorithms fail to label unanimously. \u003cem\u003eAppl Radiat Isot\u003c/em\u003e \u003cstrong\u003e176\u003c/strong\u003e, 109819 (2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e20.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Belmonte C\u003cem\u003e, et al.\u003c/em\u003e TFOS DEWS II pain and sensation report. \u003cem\u003eOcul Surf\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 404-437 (2017).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e21.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Ucar IC, Esen F, Turhan SA, Oguz H, Ulasoglu HC, Aykut V. Corneal neuropathic pain in irritable bowel syndrome: clinical findings and in vivo corneal confocal microscopy. \u003cem\u003eGraefes Arch Clin Exp Ophthalmol\u003c/em\u003e, \u0026nbsp;(2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e22.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Kobayashi A, Yoshita T, Sugiyama K. In vivo findings of the bulbar/palpebral conjunctiva and presumed meibomian glands by laser scanning confocal microscopy. \u003cem\u003eCornea\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 985-988 (2005).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e23.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;F F\u003cem\u003e, et al.\u003c/em\u003e Facial Seborrheic Keratosis with Unusual Dermoscopic Patterns can be differentiated from other skin malignancies by In Vivo Reflectance Confocal Microscopy. \u003cem\u003eJournal of the European Academy of Dermatology and Venereology : JEADV\u003c/em\u003e, \u0026nbsp;(2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e24.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Shengnan C, Yueqi Y, Jin C, Lin Y, Xinghua W, Fagang J. In vivo confocal microscopy assessment of meibomian glands microstructure in patients with Graves\u0026apos; orbitopathy. \u003cem\u003eBMC ophthalmology\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, \u0026nbsp;(2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e25.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Szalai E, Szucs G, Szamosi S, Aszalos Z, Afra I, Kemeny-Beke A. An in vivo confocal microscopy study of corneal changes in patients with systemic sclerosis. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 11111 (2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e26.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Wang J, Yeh TN, Chakraborty R, Yu SX, Lin MC. A Deep Learning Approach for Meibomian Gland Atrophy Evaluation in Meibography Images. \u003cem\u003eTransl Vis Sci Technol\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 37 (2019).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e27.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Koh YW, Celik T, Lee HK, Petznick A, Tong L. Detection of meibomian glands and classification of meibography images. \u003cem\u003eJ Biomed Opt\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 086008 (2012).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e28.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Maruoka S\u003cem\u003e, et al.\u003c/em\u003e Deep Neural Network-Based Method for Detecting Obstructive Meibomian Gland Dysfunction With in Vivo Laser Confocal Microscopy. \u003cem\u003eCornea\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 720-725 (2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e29.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Zhongwen L\u003cem\u003e, et al.\u003c/em\u003e Preventing corneal blindness caused by keratitis using artificial intelligence. \u003cem\u003eNature communications\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, \u0026nbsp;(2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e30.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Zhang K\u003cem\u003e, et al.\u003c/em\u003e Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. \u003cem\u003eNat Biomed Eng\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 533-545 (2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e31.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Yang HK, Kim YJ, Sung JY, Kim DH, Kim KG, Hwang J-M. Efficacy for Differentiating Nonglaucomatous Versus Glaucomatous Optic Neuropathy Using Deep Learning Systems. \u003cem\u003eAmerican Journal of Ophthalmology\u003c/em\u003e \u003cstrong\u003e216\u003c/strong\u003e, \u0026nbsp;(2020).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e32.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Khan SMA, Jens H, Stefan S, E SM, Philipp S. Deep learning-based automatic meibomian gland segmentation and morphology assessment in infrared meibography. \u003cem\u003eScientific reports\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, \u0026nbsp;(2021).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e33.\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Daisuke N\u003cem\u003e, et al.\u003c/em\u003e Automated detection of a nonperfusion area caused by retinal vein occlusion in optical coherence tomography angiography images using deep learning. \u003cem\u003ePloS one\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, \u0026nbsp;(2019).\u003c/p\u003e"},{"header":"Tables","content":"\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eTable 1 The total number of six types of MGD, training set ,internal and external validation set cohorts\u003c/span\u003e\u003c/p\u003e\n\u003cdiv align=\"center\" style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\n \u003ctable style=\"width:415.0pt;border-collapse:collapse;border:none;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eClassification\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eALL (n=12630)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eTrain set (n=8841)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eInternal test set (n=3789)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 81.85pt;border-color: windowtext currentcolor;border-style: solid none;border-width: 1pt medium;padding: 0in 5.4pt;height: 22.6pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eExternal test set (n=12289)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eNormal group\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2896 (22.9%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e1993 (22.5%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e903 (23.8%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:81.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2912 (23.7%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eNMGO group\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e830 (6.6%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e578 (6.5%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e252 (6.7%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:81.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e838 (6.8%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGA group\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e3585 (28.4%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2500 (28.3%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e1085 (28.6%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:81.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e3493 (28.4%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGAO group\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e1745 (13.8%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e1228 (13.9%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e517 (13.6%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:81.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e1693 (13.8%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGO group\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e3086 (24.4%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2191 (24.8%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e895 (23.6%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:81.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:22.6pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2866 (23.3%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:94.2pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:23.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGOO group\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:83.9pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:23.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e488 (3.9%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:79.85pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:23.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e351 (4.0%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.2pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:23.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e137 (3.6%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:81.85pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:23.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e487 (4.0%)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cstrong\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eAbbreviations:\u003c/span\u003e\u003c/strong\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;NMGO: normal with meibomian gland opening; MGA: meibomian gland atrophy; MGAO: meibomian gland atrophy with obstruction; MGO: meibomian gland obstruction; MGOO: meibomian gland obstruction with opening\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eTable 2 \u0026nbsp;Performance metrics of the ensemble DCNN model evaluated on the internal validation set\u003c/span\u003e\u003c/p\u003e\n\u003cdiv align=\"center\" style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\n \u003ctable style=\"width: 5.9e+2pt;background:white;border-collapse:collapse;border:none;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width:113.95pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:14.8pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003ePerformance metrics\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"6\" style=\"width:475.7pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:14.8pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eInternal test set (n=3789)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:84.85pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:11.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eNormal group (n=903)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:11.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eNMGO group (n=252)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:11.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eMGA group (n=1085)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:11.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eMGAO group (n=517)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:11.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eMGO group (n=895)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:11.3pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eMGOO group (n=137)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eAccuracy \u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.893\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.883 - 0.903)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.939\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.931 - 0.947)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.883\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.872 - 0.893)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.914\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.905 - 0.923)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.881\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.871 - 0.892)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.933\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.925 - 0.941)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eSensitivity\u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.893\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.871 - 0.912)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.889\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.843 - 0.925)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.934\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.917 - 0.948)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.919\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.892 - 0.941)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.874\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.850 - 0.895)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.876\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.809 - 0.926)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eSpecificity \u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.893\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.881 - 0.904)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.943\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.935 - 0.950)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.862\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.849 - 0.875)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.914\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.903 - 0.923)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.884\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.872 - 0.895)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.936\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.927 - 0.943)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003ePrecision\u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.723\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.696 - 0.749)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.526\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.477 - 0.574)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.731\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.707 - 0.755)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.627\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.591 - 0.661)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.699\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.672 - 0.726)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.338\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.289 - 0.390)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eNegative predictive value\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.964\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.956 - 0.970)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.992\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.988 - 0.994)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.970\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.962 - 0.976)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.986\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.981 - 0.990)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.958\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.949 - 0.965)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:18.1pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.995\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.992 - 0.997)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eKappa\u003csup\u003e\u0026sect;\u003c/sup\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.727\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.630\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.735\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.696\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.698\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.460\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eF\u003csub\u003e1\u003c/sub\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.799\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.661\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.820\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.745\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.777\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.488\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:113.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003eAUROC\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.85pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.957\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.951 - 0.964)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.6pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.965\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.954 - 0.976)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:84.25pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.959\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.954 - 0.965)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.15pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.971\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.965 - 0.977)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:75.85pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.947\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.939 - 0.955)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:76.0pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:10.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e0.963\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;color:black;'\u003e(0.947 - 0.978)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cstrong\u003e\u003cspan style='font-size: 16px; font-family: \"Times New Roman\", serif; color: rgb(0, 0, 0);'\u003eAbbreviations:\u003c/span\u003e\u003c/strong\u003e\u003cspan style='font-size: 16px; font-family: \"Times New Roman\", serif; color: rgb(0, 0, 0);'\u003e\u0026nbsp;NMGO: normal with meibomian gland opening; MGA: meibomian gland atrophy; MGAO: meibomian gland atrophy with obstruction; MGO: meibomian gland obstruction; MGOO: meibomian gland obstruction with opening;AUROC: Area Under the Receiver Operating Characteristic curve\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eTable 3 \u0026nbsp;Performance metrics of the ensemble DCNN model evaluated on the external validation set\u003c/span\u003e\u003c/p\u003e\n\u003cdiv align=\"center\" style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\n \u003ctable style=\"width: 5.9e+2pt;border-collapse:collapse;border:none;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width:92.15pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:17.35pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003ePerformance metrics\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"7\" style=\"width:496.1pt;border-top:solid windowtext 1.0pt;border-left:none;border-bottom:solid windowtext 1.0pt;border-right:none;padding:0in 5.4pt 0in 5.4pt;height:17.35pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family: \"Times New Roman\",serif;'\u003eExternal\u003c/span\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;test set (n=12289)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:13.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eNormal group (n=2912)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:13.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eNMGO group (n=838)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:13.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGA group\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(n=3493)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:13.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGAO group (n=1693)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:13.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGO group (n=2866)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:13.25pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMGOO group (n=487)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eAccuracy \u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.894\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.889 - 0.900)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.929\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.925 - 0.934)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.885\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.879 - 0.890)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.916\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.911 - 0.921)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.881\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.875 - 0.886)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.939\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.934 - 0.943)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eSensitivity\u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.901\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.890 - 0.912)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.906\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.884 - 0.925)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.940\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.931 - 0.948)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.926\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.912 - 0.938)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.893\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.881 - 0.904)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.825\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.789 - 0.858)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eSpecificity \u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.892\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.886 - 0.898)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.931\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.926 - 0.936)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.863\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.856 - 0.870)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.914\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.909 - 0.920)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.877\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.870 - 0.884)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.943\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.939 - 0.947)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003ePrecision\u003cbr\u003e\u0026nbsp;(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.722\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.707 - 0.736)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.490\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.465 - 0.515)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.732\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.718 - 0.744)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.633\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.614 - 0.652)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.689\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.673 - 0.703)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.375\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.346 - 0.405)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eNegative predictive value\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.967\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.963 - 0.970)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.993\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.991 - 0.994)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.973\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.969 - 0.977)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.987\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.985 - 0.989)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.964\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.960 - 0.968)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:21.2pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.992\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.991 - 0.994)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eKappa\u003csup\u003e\u0026sect;\u003c/sup\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.731\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.601\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.739\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.703\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.698\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.488\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eF\u003csub\u003e1\u003c/sub\u003e\u003csup\u003e\u0026dagger;\u003c/sup\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.802\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.636\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.823\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.752\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.777\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.516\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:92.15pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eAUROC\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(95% CI)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.960\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.957 - 0.963)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width:85.05pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.971\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.966 - 0.976)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:78.0pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.961\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.958 - 0.964)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:99.2pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.974\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.971 - 0.977)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.954\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.950 - 0.958)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:77.95pt;border:none;border-bottom:solid windowtext 1.0pt;padding:0in 5.4pt 0in 5.4pt;height:12.0pt;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e0.951\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e(0.941 - 0.961)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"border:none;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cstrong\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eAbbreviations:\u003c/span\u003e\u003c/strong\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;NMGO: normal with meibomian gland opening; MGA: meibomian gland atrophy; MGAO: meibomian gland atrophy with obstruction; MGO: meibomian gland obstruction; MGOO: meibomian gland obstruction with opening; AUROC: Area Under the Receiver Operating Characteristic curve\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n\u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eTable 4 Application of deep learning in ophthalmic diseases\u003c/span\u003e\u003c/p\u003e\n\u003ctable style=\"border-collapse:collapse;border:none;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78pt;border-color: windowtext currentcolor;border-style: solid none;border-width: 1pt medium;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eAuthor\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 42.5pt;border-color: windowtext currentcolor;border-style: solid none;border-width: 1pt medium;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eYear\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 121.1pt;border-color: windowtext currentcolor;border-style: solid none;border-width: 1pt medium;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eDiseases\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 119.9pt;border-color: windowtext currentcolor;border-style: solid none;border-width: 1pt medium;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eDeep learning algorithms\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53pt;border-color: windowtext currentcolor;border-style: solid none;border-width: 1pt medium;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eRefs.\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eLi \u003cem\u003eet al\u003c/em\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 42.5pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2021\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 121.1pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eKeratitis\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 119.9pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eDenseNet121,\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eInception-v3,\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eResNet-50\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003csup\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e29\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eZhang \u003cem\u003eet al\u003c/em\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 42.5pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2021\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 121.1pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eChronic kidney disease and diabetic fundus disease\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 119.9pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eResNet-50\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003csup\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e30\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eYang \u003cem\u003eet al\u003c/em\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 42.5pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2020\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 121.1pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eGlaucomatous Optic Neuropathy\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 119.9pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eResNet-50\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003csup\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e31\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eKhan \u003cem\u003eet al\u003c/em\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 42.5pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2021\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 121.1pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eMeibomian gland dysfunction\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 119.9pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eResNet-v2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53pt;border: medium none;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003csup\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e32\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78pt;border-color: currentcolor currentcolor windowtext;border-style: none none solid;border-width: medium medium 1pt;border-image: none 100% / 1 / 0 stretch;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eDaisuke \u003cem\u003eet al\u003c/em\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 42.5pt;border-color: currentcolor currentcolor windowtext;border-style: none none solid;border-width: medium medium 1pt;border-image: none 100% / 1 / 0 stretch;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e2019\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 121.1pt;border-color: currentcolor currentcolor windowtext;border-style: none none solid;border-width: medium medium 1pt;border-image: none 100% / 1 / 0 stretch;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eBranch Retinal Vein Occlusion\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 119.9pt;border-color: currentcolor currentcolor windowtext;border-style: none none solid;border-width: medium medium 1pt;border-image: none 100% / 1 / 0 stretch;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003eVisual Geometry Group (VGG)-16\u003c/span\u003e\u003c/p\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 53pt;border-color: currentcolor currentcolor windowtext;border-style: none none solid;border-width: medium medium 1pt;border-image: none 100% / 1 / 0 stretch;padding: 0in 5.4pt;vertical-align: top;\"\u003e\n \u003cp style=\"margin:0in;text-align:justify;font-size:14px;font-family:DengXian;\"\u003e\u003csup\u003e\u003cspan style='font-size:16px;font-family:\"Times New Roman\",serif;'\u003e33\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFoundation item\u003c/strong\u003e\u003cstrong\u003e：\u003c/strong\u003eNational\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eNatural Science Foundation（No:\u0026nbsp;82160195）; \u0026nbsp;Central Government Guides Local Science and Technology Development\u0026nbsp;Foundation(No:\u0026nbsp;20211ZDG02003);\u0026nbsp;Key Research Foundation of Jiangxi Province (No: 20181BBG70004, 20203BBG73059); Excellent Talents Development Project of Jiangxi Province（No:\u0026nbsp;20192BCBL23020）; Natural Science Foundation of Jiangxi Province（No:\u0026nbsp;20181BAB205034）; Grassroots Health Appropriate Technology \u0026ldquo;Spark Promotion Plan\u0026rdquo; Project of Jiangxi Province(No:20188003); Health Development Planning Commission Science Foundation of Jiangxi Province (No: 20201032,202130210)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analyzed during the present study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study methods and protocols were approved by the Medical Ethics Committee of the First Affiliated Hospital of Nanchang University (Nanchang, China) and followed the principles of the Declaration of Helsinki. All subjects were notified of the objectives and content of the study and latent risks, and then provided written informed consent to participate.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePatient consent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study did not receive any industrial support. The authors have no competing interests to declare regarding this study\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003eAll the authors contributed to this manuscript.\u0026nbsp;Yi Shao\u0026nbsp;and Yi-Chen Yang\u003cem\u003e\u0026nbsp;\u003c/em\u003eare responsible for conceiving and designing the work, acquiring data and writing the manuscript; Hui Zhao and Wen-Qing Shi\u003cem\u003e\u0026nbsp;\u003c/em\u003eplayed an important role in interpreting the results; Xu-Lin Liao, Ting Su and Rong-Bin Liang Min\u003cem\u003e\u0026nbsp;\u003c/em\u003ehelped in acquiring data and giving some advice; Qiu-Yu Li and Qian-Min Ge\u003cem\u003e\u0026nbsp;\u003c/em\u003econtributed to the application to the ethics committee; Hui-Ye Su and Yi-Cong Pan revised the manuscript, and Xiang-Chun Li gave valuable guidance at every stage and approved the final version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"meibomian glands, in vivo confocal microscopy, deep learning","lastPublishedDoi":"10.21203/rs.3.rs-936418/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-936418/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"In vivo confocal microscopy（IVCM） is a real-time in vivo high-resolution and non-invasive imaging method that allows observation of morphological changes in the lid gland at the cellular level. We use IVCM to observe the meibomian glands and divide the 12,630 pictures obtained into six groups (normal group, normal with meibomian gland opening group, meibomian gland atrophy group, meibomian gland atrophy with obstruction group, meibomian gland obstruction group and meibomian gland obstruction with opening group), randomly select 70% of the pictures and use the ResNet34 deep learning network model for training, and use the remaining 30% of the pictures and another 12889 pictures collected from the top three hospitals as internal and external validation sets to verify the performance of the model. The results show that the model validation set recognition constructed using the deep learning method has achieved good performance, and the obtained AUROCs are all greater than 0.95. This model provides the possibility for future automatic classification and diagnosis of meibomian gland dysfuncton(MGD),and can be used for MGD related clinical auxiliary diagnosis and screening of diseases.","manuscriptTitle":"Detection of Meibomian Gland Dysfunction by in vivo Confocal Microscopy Based on Deep Convolutional Neural Network","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2021-10-25 17:17:52","doi":"10.21203/rs.3.rs-936418/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"61d5e642-d0d4-4372-b425-d33ef779ee03","owner":[],"postedDate":"October 25th, 2021","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":8092939,"name":"Computational Biology"},{"id":8092940,"name":"Bioinformatics"}],"tags":[],"updatedAt":"2024-02-19T10:45:29+00:00","versionOfRecord":[],"versionCreatedAt":"2021-10-25 17:17:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-936418","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-936418","identity":"rs-936418","version":["v1"]},"buildId":"7rjqhiLT3MXkJMwkYKINL","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}