Initiation of China Alliance of Research in High Myopia (CHARM): protocol for an AI-based multimodal high myopia research biobank


Myopia, particularly high myopia, is a growing public health concern globally, affecting millions of people worldwide. The estimated prevalence of myopia and high myopia is projected to increase from 1460 million and 163 million in 2000 to 4758 million and 938 million in 2050, respectively.1 The high prevalence and incidence of myopia impose a significant socioeconomic burden due to the loss of productivity.2 East Asian countries, especially China, have the highest prevalence of myopia, with studies indicating that more than 50% of high school students and young adults are affected by the condition.3 Furthermore, the increasing prevalence of myopia among younger children and higher rates of progression during school years have led to an escalating burden.4

High myopia is defined as a high degree of myopic refractive error ≤−6.00D and/or axial length (AL) ≥26.0 mm,5 while pathological myopia is characterised by myopic maculopathy equal to or more severe than diffuse choroidal atrophy (META-analysis of Pathologic Myopia, META-PM) or the presence of posterior staphyloma.6 High myopia and pathological myopia are major causes of visual impairment. High myopia is associated with an increased risk of myopic macular degeneration (OR=845.08), retinal detachment (OR=12.62), posterior subcapsular cataract (OR=4.55) and open-angle glaucoma (OR=2.92).7 Therefore, identifying myopia and its progression to high myopia is crucial, as it is the primary and only modifiable risk factor for pathological myopia.8

Early detection and diagnosis of myopia are crucial in children for managing its progression and reducing the risk of myopic maculopathy, making myopia screening an urgent demand. However, the identification of the ocular fundus that requires treatment or scheduling a follow-up visit is a time-consuming and labor-intensive task that requires specialised expertise. The advancement of artificial intelligence (AI) can aid in the grading of myopic maculopathy and perform as well for general ophthalmologists and retinal specialists.9 In addition to its success in detecting high myopia and pathological myopia, AI has demonstrated high sensitivity and specificity in identifying diabetic retinopathy and related eye diseases using retinal images from individuals with diabetes.10 Moreover, a deep learning algorithm based on retinal photography has the potential to yield significant insights into cerebrovascular neurodegenerative processes, and potentially plays a role in studying dementia, stroke and Alzheimer’s disease, offering a complementary form of neuroimaging.11 12

To address this challenge, a high-quality big data platform is necessary. We have launched a major research initiative called the China Alliance of Research in High Myopia (CHARM) supported by AI technology to analyse the vast amounts of data collected from study participants. CHARM is an AI-assisted high myopia project with more than 100 hospitals in China participating; the geographical distribution of the institutions that are currently participating is shown in figure 1. In the future, new institutions will continue to join the alliance. The project aims to screen, diagnose and treat high myopia patients and serves as an excellent platform to facilitate collaboration beyond preventive ophthalmology in areas where big data is required.

Figure 1
Figure 1

Geographical distribution of member units. Different dots represent the units in which data have been uploaded. In the future, there will be new hospitals and institutions continuing to join the alliance.

This article outlines the design, inclusion and exclusion criteria, task assignment, data upload, cleansing and analysis, intellectual property rights, and quality assurance and confidentiality clause of CHARM.


Research on the topic of ‘AI and myopia’

On 1 May 2023, a comprehensive literature search was conducted on publicly available databases including PubMed, Web of Science and Medline, without any restrictions on the start date. The details of search strategy were (“myopia”[MeSH Terms] OR “myopia”[All Fields] OR “myopias”[All Fields] OR (“high”[All Fields] AND (“myopia”[MeSH Terms] OR “myopia”[All Fields] OR “myopias”[All Fields])) OR ((“pathologic”[All Fields] OR “pathologically”[All Fields] OR “pathologics”[All Fields] OR “pathology”[MeSH Terms] OR “pathology”[All Fields] OR “pathological”[All Fields]) AND (“myopia”[MeSH Terms] OR “myopia”[All Fields] OR “myopias”[All Fields]))) AND (“artificial intelligence”[MeSH Terms] OR (“artificial”[All Fields] AND “intelligence”[All Fields]) OR “artificial intelligence”[All Fields] OR (“antagonists and inhibitors”[MeSH Subheading] OR (“antagonists”[All Fields] AND “inhibitors”[All Fields]) OR “antagonists and inhibitors”[All Fields] OR “ai”[All Fields]) OR (“machine learning”[MeSH Terms] OR (“machine”[All Fields] AND “learning”[All Fields]) OR “machine learning”[All Fields]) OR (“deep learning”[MeSH Terms] OR (“deep”[All Fields] AND “learning”[All Fields]) OR “deep learning”[All Fields])).

A database search on the topic of AI and myopia yielded nearly 1000 records. Previous studies in this field have primarily focused on two areas: (1) the use of ophthalmic images for disease diagnosis through validation and evaluation of DL models, including the screening of high myopia using spectral domain optical coherence tomography (OCT),13 classification of different levels of myopic maculopathy using fundus photographs sourced from clinics,14 development of a machine learning approach for automatic detection of pathological myopia based on fundus images15 and evaluation of the clinical usefulness of novel denoising processes for depicting myopic choroidal neovascularisation16 and (2) determining prognosis from these images, such as identifying children at risk of developing high myopia for timely assessment and intervention, and preventing myopia progression and complications in adulthood through the development of a DL system.17

Tan and colleagues conducted a retrospective multicohort study in which they developed and tested DL algorithms for the detection of myopic macular degeneration and high myopia, using 226 686 retinal images from datasets in Singapore. The use of blockchain technology could potentially provide a trusted platform for future AI model performance testing in medicine.18 However, there has not yet been a comprehensive multicentre study to evaluate its efficacy and ability to detect and quantify fundus photograph parameters undertaken in China. Therefore, the development of CHARM is imperative for assessing and optimising AI algorithms for fundus photograph analysis in the Chinese population.


CHARM is a collaborative research project between multiple hospitals, institutions and universities in China led by the Beijing Institute of Ophthalmology/Beijing Tongren Hospital, Capital Medical University. The study has received approval from the Ethics Committee of Beijing Tongren Hospital (TREC2022-KY045), and all methods are in accordance with the Declaration of Helsinki. CHARM includes retrospective analysis of existing data and follow-up studies of subsequent data. Informed consent was waived for the retrospective study, while informed consent was obtained for the cohort study. Before the study begins, a tripartite contract must be signed among three parties, including the organiser (Party A: Beijing Institute of Ophthalmology), participant hospitals or academic institutions (Party B) and the operator (Party C: EVision Technology (Beijing) Co), to clarify their tasks and rights. The cooperation and research can only be implemented after all three parties have no objection to the contract. Z-BJ, Director of the Beijing Institute of Ophthalmology, proposed the CHARM module. The core data storage and analysis were supported by EVision Technology. Additional support for training, implementation and quality control of CHARM came from the advisory board; the management mode of CHARM is shown in figure 2.

Figure 2
Figure 2

Current operation mode. The overall management is carried out by the executive chairman, guided by an advisory board composed of three renowned experts, and supported by the secretariat and the technical support team. The leaders of the five main units, in terms of the number of data uploaded, are selected to be members of the committee that jointly participates in the management of the China Alliance of Research in High Myopia.

The study is carried out in a three-step strategy, including the collection of basic information and multimodal imaging data, and the implementation of genetic testing based on the patient’s family history and blood samples. The ‘first-step’ step: eye data available within CHARM include single-field fundus images (45° centred between the optic disc and the macula), AL and refraction (in dioptre (D)) by age, gender and ethnic data of more than 100 000 people. In addition, ultra-widefield fundus images, macular OCT as well as other imaging examinations will be carried out in the ‘second-step’ step. Other data will be available in the ‘third-step’ step, together with the longitudinal design of the study, making this possibly the most valuable research resource for CHARM.

The study aims to develop new tools and technologies to detect and diagnose myopic maculopathy, improve our understanding of myopia and develop effective strategies to prevent it, identify the genetic and environmental factors that contribute to the development of high myopia and make a novel classification method of high myopia based on fundus images by AI.

Inclusion and exclusion of the first-step step

The first-step step of the study includes patients diagnosed with high myopia with no age limit, with a spherical equivalent (SE) of ≤−6.00 D and/or AL of ≥26.0 mm. Demographic and clinical characteristics, such as age, gender and medical history (including data related to myopia progression reduction therapies like low-dose atropine, multisegment glasses and multifocal contact lenses), should be documented. Additionally, it is advisable to record information about ethnic groups, with special attention to ethnic minorities.

Fundus photography: High-quality 45° images covering the entire fovea, optic nerve head, and upper and lower vascular arches, which can be analysed by AI were enrolled, which should be submitted in jpg format (Digital Imaging and Communications in Medicine (DICOM) is also acceptable).

Refraction status: Subjective refraction, retinoscopy, autorefraction or cycloplegic autorefraction were all preferred as an alternative. For children under 12 years, it is recommended to use the results of refraction after cycloplegia as the final outcome. Additionally, for those under 16 years, performing cycloplegic refraction is recommended but not mandatory. All measurements (spherical, cylindrical and axis) should be recorded. Subsequently, The SE of the refractive error was calculated as spherical refractive error+1/2 cylindrical refractive error. The method of measurement used must also be noted.

Axial length: AL should be measured by ultrasonic biometry and optical biometry, such as contact applanation A-scan ultrasonography, IOLMaster or some new types of optical biometer. The method of measurement used needs to be noted.

The presence of ocular and systemic diseases in patients may have an effect on the image processing of fundus photographs. Therefore, patients with conditions such as all types of glaucoma, diabetic retinopathy, retinal vein occlusion, anterior ischaemic optic neuropathy and neurological diseases, as well as those with a history of intraocular surgery, were excluded from the study. Incomplete data were also excluded. It is important to note that fundus and refractive parameters before intraocular collamer lens surgery and posterior scleral reinforcement were included in the study, as these morphological characteristics remain unchanged.

Task assignment

CHARM was initiated by the Beijing Institute of Ophthalmology (Party A), together with member units (Party B) and technical support unit EVision Technology (Beijing) Co (Party C). Party A will establish CHARM, formulate data collection standards, organise data cleaning and labelling, and coordinate the overall scheme. Party B will join CHARM for data collection and uploading, while Party C will assist in building a public dataset and be responsible for the quantitative processing of fundus images.

The specific responsibilities of the parties are as follows:

  1. Party A will develop public dataset standards, and organise data collection, cleaning, annotation and processing by Parties B and C according to the standards.

  2. Party B will collect data according to Party A’s standards, provide effective data for at least 200 patients and participate in data cleaning, annotation and processing.

  3. Party C will assist Party A in establishing the alliance and provide technical support for data collection, cleaning, annotation and processing, such as building the data platform.

  4. The parties will conduct scientific research based on the dataset, report research achievements to other units within 1 month and prioritise research ideas proposed by Party B based on the equity coefficient.

  5. Each unit should designate a responsible person to carry out relevant work and notify other units of any changes in personnel.

Intellectual property rights

To better run the alliance, the intellectual property rights are formulated as follows:

  1. All parties have the right to use the data after the dataset is established, and the ownership of the data provided by each unit still belongs to their own institutions.

  2. The dataset is confidential and should not be provided to non-alliance units for use. Intellectual property rights are jointly held by all participating parties, and independently generated intellectual property rights belong to the producing unit.

  3. The unit that contributes more than 10% can be designated as the corresponding author for final publications.

  4. Authorship order will be negotiated based on actual contributions made by each member unit, and equity coefficients based on contributions to the effective dataset will be set. Public dataset use should be clearly indicated, and research conclusions should avoid conflicts.

The data upload process

To make data uploading and management more convenient, the technical support unit has developed a cloud data management service platform that facilitates the uploading of data from various alliance units and the management of alliance data. Participating units can upload data individually or in batches after organising them on the cloud platform. The platform will perform deidentification to safeguard sensitive information contained in the fundus photographs from participants (eg, names or identification numbers) in uploaded data; our fundus image intelligent analysis software—EVisionAI—will be employed to automatically redact these details,19 20 effectively concealing or removing them from the image, and other units except for the uploader will not be able to see the sensitive information. Additionally, the platform will perform uniform structured processing on the uploaded data and automatically check and remind of any obvious content anomalies, thus preventing errors caused by information entry. Finally, all data uploaded to the cloud platform will be stored in encrypted form to ensure data security.

Data cleaning and fundus images analysis

The technical support unit will conduct cleaning and analysis for each uploaded datum according to the alliance’s requirements. The cleaning process includes two parts: evaluation of retinal image quality and completeness of data.

Retinal image quality evaluation: Starting from the image content itself, the optic disc and macula are automatically identified and located on the image based on computer vision technology.21 The exposure and clarity of the image are automatically calculated. Based on the identification results of the optic disc and macula in the retinal image, the eye type (left/right) and eye position (optic disc or macula centre) of the image are automatically determined. Furthermore, by combining the results of the exposure and clarity calculation, the image is judged whether it is qualified, and unqualified images are removed.

Data completeness organisation: After the evaluation of retinal image quality, the various dimensions of data are comprehensively checked. Data with incomplete dimensions are directly removed, and the remaining data are organised and grouped according to eye type and eye position.

After completing data cleaning, we employ the fundus image intelligent analysis software EVisionAI for intelligent image processing. This software is designed based on the biomimetic principles of human vision and seamlessly integrates advanced image processing technologies like computer vision and DL. The technical support unit adopts the human visual biomimetic approach and uses AI technologies, including computer vision and DL, to conduct unified quantitative processing on retinal images. First, a series of preprocessing operations including region of interest extraction, denoising, normalisation and enhancement is conducted on all images to improve the sharpness of internal image features and reduce interimage differences. Next, the unit comprehensively combines colour, brightness, shape and other information displayed within retinal features, deeply integrating the DL network model with the computer vision algorithm based on the visual attention mechanism (figure 3).22–24

Figure 3
Figure 3

Fundus photo processing with artificial intelligence deep learning system. Extraction of optic disc and PPA, central retinal artery and vein, and fundus tessellation. PPA, peripapillary atrophy.

Optic disc extraction: Initially, the optic disc is identified using a DL object detection model known as single shot detection with ResNet-50 as the backbone network. The model is trained on training samples to obtain the bounding box of the optic disc, with its centre point serving as the optic disc centre. Subsequently, a visual attention mechanism is employed to determine the optic disc’s boundary with its centre as the origin. The fundus image is then transformed into polar coordinates, and an edge detection operator is applied to obtain the polar coordinates’ optic disc edge. Finally, the image is inverse-transformed into rectangular coordinates, achieving fine segmentation of the optic disc. A smallest circumscribed circle fitting is performed on the segmented optic disc area to determine its final centre point and diameter.

Peripapillary atrophy (PPA) extraction: The preprocessed image is fed into a segmentation model, which uses the ResNet101-UNet network for PPA arc segmentation. The model is trained on sample data from the training set and evaluated using data from the verification set. The test set is used to assess segmentation accuracy, calculating metrics like accuracy, sensitivity, specificity and intersection over union. The optimal PPA segmentation model is trained, allowing for PPA region extraction. PPA width is defined as the longest distance between the intersection points of the centreline and the PPA area’s edges.

Vessel extraction: The preprocessed fundus image undergoes vessel segmentation using the DL network ResNet101-UNet. Arterial and venous vessels are distinguished based on colour, brightness features, vessel connectivity and topological relationships. Morphological erosion is applied to the segmented vessels to obtain vessel centrelines. The arteriovenous diameter ratio is determined as the diameter ratio between arterial and venous vessels. Vessel tortuosity is computed as the curvature along the vessel centreline. The average curvature is calculated for the vessel centreline. The optic disc’s diameter is defined based on the diameter of the smallest circumscribed circle of the segmented optic disc area. Various vascular parameters are measured, including vascular diameter, vascular density and branching angles. Vascular arch parameters are derived from the centre of the optic disc and aortic veins. Vascular density is the ratio of retinal vascular area to the fundus area. Vascular fractal dimension measures the complexity of retinal vascular branches. Venular length is the sum of venule centreline lengths.

It recognises and segments the features of the optic disc, blood vessels, PPA and exposed choroidal vessels. Based on the segmentation results and using the optic disc centre or macula centre as the reference, it measures and quantifies various structural features such as the diameter/area of the optic disc, diameter/curvature of the blood vessel, width/height of PPA and fundus tessellated density. It uses multidimensional digital indicators to comprehensively, precisely and objectively describe the features of the retinal structure, which can assist in subsequent studies on the mechanisms of myopia formation and progression.

Quality assurance and confidentiality clause

To ensure the successful implementation of CHARM, a quality assurance programme will be established through rigorous methodological validation established by the advisory board. The ophthalmologists will provide uniform and specialised operational training to all staff who will participate in data collection procedures, including fundus photo taking, refraction status assessment, AL measurement and demographic data investigation, to ensure the standardisation and high quality of the uploaded data. EVision Technology will send technicians to participating medical institutions and hospitals to conduct fundus scans, and AI experts will analyse the images. Highly trained personnel and members will conduct training online or offline to ensure the homogeneity and quality of samples collected from different medical institutions. Due to the clear workflow and multicentre coordination, facilitates regular meetings and communication, which can ensure consistency in the protocols and procedures of CHARM.

During the collaboration, if any of the image materials provided by any party involve clinically sensitive information about patients, all parties and alliance units involved should strictly comply with medical ethics and protect patients’ privacy by masking relevant information before publication. Before the dataset is made publicly available, no party is allowed to disclose any information related to the dataset or provide the data involved in the dataset to non-alliance units. Any technical information, data or other information provided by any party during the establishment and research process of the dataset, which has not been made public or has been informed of its inability to be provided to a third party, should be kept confidential by the alliance units without the consent of the provider. This confidentiality clause will remain in effect even after the termination of the cooperation agreement.

Patient and public involvement

There is no patient or public involvement in our research’s design, conduct, reporting or dissemination plans.

Discussion and conclusion

CHARM is a collaborative project comprising multiple institutions and universities in China, with the objective of constructing the world’s largest and most comprehensive public database on myopia. The project also aims to establish a myopia research service platform that serves the global community and investigates the underlying causes and developmental mechanisms of myopia with the goal of promoting new advances and modalities in myopia control.

The project takes a three-step strategy that is closely linked and will be carried out step by step. The aims of the establishment of CHARM are as follows. First, developing new tools and technologies to detect and diagnose myopic maculopathy at an earlier stage, using AI-assisted analysis of large datasets of medical images and clinical data to accurately identify early biomarkers of myopia and predict which individuals are at the highest risk of developing the condition.

Second, identify the genetic and environmental factors that contribute to the development of high myopia in our third-step step. Researchers are collecting data from Chinese individuals with high myopia and their families to pinpoint the specific genes and environmental factors that play a role in the condition, developing more targeted interventions for preventing and treating the condition.

Third, explore new treatments for high myopia and investigate the role of novel modalities in preventing and treating myopia. The project also focuses on primary, secondary and tertiary prevention to reduce the incidence and slow the progression of myopia, which is a government priority in China.4

CHARM is a significant undertaking that has the potential to improve the lives of millions of people with high myopia in China and around the world. Researchers from multiple institutions and disciplines can pool their resources and expertise to tackle this challenging problem by working together. Through our efforts, we may have a better understanding of myopia and effective strategies to prevent and treat the condition in the near future.

Overall, CHARM aims to serve as a model for collaborative ophthalmological research, highlighting the power of interdisciplinary efforts and cutting-edge technologies in addressing complex health challenges.

This post was originally published on