Performance of artificial intelligence for the detection of pathological myopia from colour fundus images: a systematic review and meta-analysis

Background Pathological myopia (PM) is a major cause of worldwide blindness and represents a serious threat to eye health globally. Artificial intelligence (AI)-based methods are gaining traction in ophthalmology as highly sensitive and specific tools for screening and diagnosis of many eye diseases. However, there is currently a lack of high-quality evidence for their use in the diagnosis of PM. Methods A systematic review and meta-analysis of studies evaluating the diagnostic performance of AI-based tools in PM was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidance. Five electronic databases were searched, results were assessed against the inclusion criteria and a quality assessment was conducted for included studies. Model sensitivity and specificity were pooled using the DerSimonian and Laird (random-effects) model. Subgroup analysis and meta-regression were performed. Results Of 1021 citations identified, 17 studies were included in the systematic review and 11 studies, evaluating 165,787 eyes, were included in the meta-analysis. The area under the summary receiver operator curve (SROC) was 0.9905. The pooled sensitivity was 95.9% [95.5%-96.2%], and the overall pooled specificity was 96.5% [96.3%-96.6%]. The pooled diagnostic odds ratio (DOR) for detection of PM was 841.26 [418.37–1691.61]. Conclusions This systematic review and meta-analysis provides robust early evidence that AI-based, particularly deep-learning based, diagnostic tools are a highly specific and sensitive modality for the detection of PM. There is potential for such tools to be incorporated into ophthalmic public health screening programmes, particularly in resource-poor areas with a substantial prevalence of high myopia.


INTRODUCTION
Myopia is one of the most common ocular conditions worldwide, with global prevalence predicted to increase from nearly 2.8 billion in the year 2020 to almost 5 billion-~49.8% of the world's population-by the year 2050 [1].High myopia, generally defined as a refractive error of −6 dioptres (D) or greater, can predispose individuals to sight-threatening sequelae such as glaucoma, cataract, retinal tears or detachment.
Pathological myopia (PM)-which occurs as a result of structural changes in the posterior segment of the eye due to significant axial elongation [2], is one of the major causes of irreversible visual impairment worldwide [2][3][4][5], affecting ~3% of the world population and as many as 50-70% of high myopics to some degree [6].Reduced visual acuity due to PM can result in a considerable negative impact on quality of life, including social and emotional health and functional ability [7].The potential economic impact of PM is also profound; a 2015 metaanalysis estimated the global productivity loss caused by myopic macular degeneration to be around US $6 billion worldwide [8].
The prevalence of myopia-the main risk factor for PM development-is extreme in many areas; in one study of 23,616 males in South Korea, 96.5% were myopic [9].Evidence suggests that treatment failure in the correction of myopia is common and that long-term efficacy (of importance in reducing the risk of PM) is often limited [10].As a result, a significant number of individuals, particularly in highly myopic populations, are still likely to develop PM, underscoring the need for cost-effective, reliable and scalable screening programmes to identify and monitor patients with PM and follow-up those at high risk of developing sight-threatening complications.
The diagnosis of PM is made qualitatively on fundal examination.Qualitative diagnosis can be subject to inter-observer variability between practitioners, and requires considerable clinical expertise to perform accurately.While optical coherence tomography (OCT) may also be used in the diagnosis of PM, retinal fundus photography remains the most widely accessible form of ophthalmic imaging for screening purposes.Hence, at present, retinal fundus images are likely to be the most useful modality to test the efficacy of new diagnostic aids and tools.
Several 'classical' features of PM may be visualised on fundus imaging, including a 'tesselated' atrophy of the retinal pigment epithelium, peripapillary atrophy, temporal flattening of the optic disc, lacquer cracks, posterior staphyloma and Fuch's spot.A common cause of blindness in PM is myopic choroidal neovascularisation (CNV), which carries an extremely poor prognosis if untreated [2,11].
Artificial intelligence (AI)-based diagnostic tools seek to reduce the need for expert interpretation by learning the features of normal and abnormal examples, with the aim of being able to label images autonomously.AI-aided diagnosis is no longer a novel concept in ophthalmology, and has been the subject of much evaluation for the screening of multiple ocular diseases, such as age-related macular degeneration, glaucoma, diabetic retinopathy, papilloedema and retinopathy of prematurity [12][13][14][15][16].However, no systematic review or meta-analysis to date has sought to collate and evaluate the efficacy of these methods for the diagnosis of PM.
Therefore, the aim of the present systematic review and metaanalysis is to assess the diagnostic accuracy of artificial intelligence-based methods for the detection of PM using colour fundus images.

Study registration
This study was registered on PROSPERO with registration number CRD42022309830.

Search strategy and inclusion/exclusion criteria
According to the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA), and using a search strategy designed by JP (Supplementary Table 1), the MEDLINE, EMBASE, CINAHL, Web of Science and IEEEExplore databases were searched.Reference lists of included studies were subsequently hand searched to identify additional studies that met the predefined inclusion criteria.
Studies were included if they reported the effectiveness of machine learning-or artificial intelligence-based detection algorithms in detecting PM; used indices such as area under the receiver-operator curve (AUROC), sensitivity and specificity to report on algorithm performance; evaluated colour fundus images; provided information about the size of the dataset and the reference standard; included a validation set at least 10% of the size of the training set; were in English and were published in a peerreviewed journal.Reviews and conference abstracts were not included.

Study selection
Both reviewers independently screened all citations (and subsequently the full texts of included citations) for inclusion in a blinded process.Disagreements were resolved via mutual discussion, and details of these disagreements, along with final decisions on inclusion, are included in Supplementary Table 7.

Data extraction and quality assessment
A single reviewer (NT) extracted data from the included studies (Table 1, 2).Extracted data were directly checked against study data by a second reviewer (JP).Attempts were made to contact study authors for any missing information.Risk of bias assessment was performed using a novel, multi-step approach combining the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2 [17]) checklist and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM [18]).Each reviewer performed quality assessment independently using both checklists.Areas of conflict were highlighted and are included in Supplementary Table 7.

Data analysis
2 × 2 contingency tables and statistics were generated using the best available data from study manuscripts (Table 1).Using the 2 × 2 contingency tables extracted for each study, sensitivity, specificity, positive predictive value, negative predictive value, negative likelihood ratio and positive likelihood ratio were calculated.Ungradable images were excluded from all analyses.
Where studies reported testing on both a primary, held-out test set and one or more external validation sets, the 2 × 2 contingency tables for each were combined to obtain a single contingency table.The combined contingency table was used in the meta-analysis.The contingency tables and model performance for the individual sets were recorded.
The presence of heterogeneity between included studies was assessed using the chi-square test and quantified by Higgins' I 2 [19], where significant heterogeneity was considered to be I 2 ≥ 50%.The presence of threshold effects was assessed using the Spearman correlation coefficient between the logit of the true positive rate (TPR) and the false positive rate (FPR).
Where the heterogeneity among the included studies exceeded the stated threshold of 50%, measures were pooled using the DerSimonian and Laird (random-effects) model [20].Where heterogeneity among the studies did not include the threshold, we planned to use a Mantel-Haenszel (fixed-effects) model [21].
Heterogeneity was investigated using subgroup analyses and metaregression.Subgroup analyses were pre-specified in the study protocol, and included the type of model used and country of study/dataset origin.Sensitivity analyses were performed to assess the relationship between reviewer-assessed study quality and diagnostic accuracy and heterogeneity, in line with methods described by Higgins et al. [19].Metaregression was performed to analyse the relative effect of the size of the training set in each study.
The SROC curve was used to visually describe the relationship between the TPR and FPR in the included studies.The area under the SROC curve (AUROC) was calculated to demonstrate the probability of a classifier correctly classifying a randomly selected negative and positive example.Fagan nomograms were generated to describe the pre-test (prior) and post-test (posterior) probability for included studies, enabling direct translation of our results to the clinical setting.

Study quality assessment and publication bias
The results of the quality assessment are reported in Fig. 2, and Supplementary Tables 2 and 3.The quality of included studies was fair, with all studies achieving either moderate or high quality.There were no disagreements between reviewers on the quality of included studies.No articles were excluded on the basis of poor quality.
Using the quantitative funnel plot test described by Deeks et al. [23], it was determined that publication bias was unlikely for studies included in the meta-analysis (t = −1.53,p = 0.1607).The qualitative funnel plot is shown in Supplementary Fig. 1.

Study characteristics
All 17 included studies described and evaluated an AI-based method to identify pathological myopia from colour fundus images.Fourteen studies (82.4%) used convolutional neural network-based methods, with one of these studies also using a support vector machine (SVM) method and another using a k-nearest neighbours method for classification.Two studies (11.8%) [32,38] used SVM for classification, and another [40] used joint sparse multi-task learning.Studies excluded from the meta-analysis as no contingency table could be formed from the published data. b Only AUC and Accuracy were provided for external validation sets, hence these were excluded from the meta-analysis.
Six studies (35.3%) used publicly available datasets, and two studies (11.8%) used some publicly available data.Demographic information, where reported, is presented in Supplementary Table 8.Eight studies (47.1%) used the META-PM definition for pathological myopia to guide annotation; the remainder did not use a formal definition.Eight studies (47.1%) included an external validation set.
Seven studies (41.2%) compared the performance of the algorithm with that of one or more human graders, reported in Supplementary Table 10.
Fourteen studies (82.4%) used direct labelling by expert ophthalmologists or retinal specialists only as the reference standard.One study (5.9%) also used labelling by expert ophthalmologists and non-medical expert graders.One study (5.9%) used self-labelling methods for the training data and health record data for the test set.Two studies (11.8%) used health record data to generate labels.
Individual contingency tables for included studies are reported in Supplementary Table 5.
Performance comparison with human graders 7 of 11 studies included in the meta-analysis (63.6%) also reported a comparison with human graders, and are reported in Supplementary Table 10.Where reported, the mean sensitivities and specificities of the human graders ranged from 0.719-0.986and 0.972-0.998,respectively.The corresponding proposed model sensitivities and specificities ranged from 0.908-0.991and 0.925-1.000,respectively.

Heterogeneity analysis
Heterogeneity across the included studies was substantial (Fig. 3).We sought to explain this heterogeneity using threshold analysis, meta-regression and subgroup analysis.

Threshold analysis
A threshold analysis showed no significant effect (Spearman correlation coefficient = −0.112,p = 0.729), indicating that the heterogeneity observed between the included studies was unlikely to be due to a threshold effect.

Subgroup and sensitivity analyses
We performed subgroup analyses as pre-specified in the study protocol.Subgroup analysis by publication year was deemed inappropriate due to the recent publication date of all except one study.
A subgroup analysis by classification algorithm was performed.Two studies [25,31] used an SVM-based model for classification and were considered as a separate subgroup to studies using CNNs.Fifteen studies used a CNN-based approach.Although the two non-CNN studies had a significantly lower pooled sensitivity (0.878 [95% CI 0.832-0.915],I 2 = 0.0%) than the CNN studies (0.960 [95% CI 0.957-0.964],I 2 = 97.4%), the pooled specificity was higher (0.999 [95% CI 0.998-1.000],I 2 = 94.3%)than that of the CNN studies (0.961 [95% CI 0.959-0.962],I 2 = 99.3%).The AUROC for the CNN studies was 0.9925; as there were only two non-CNN studies, AUROC was not estimable for this subgroup.Subgroup analysis by study country (studies originating from China vs. other regions) and showed lower sensitivities, specificities and AUROC for studies originating outside China (Supplementary Figs. 2 and 3).Heterogeneity generally remained high (I 2 > 90%) in subgroups, limiting interpretation.

Meta-regression
Univariate meta-regression (Table 2) showed no statistically significant effect of training set size to account for the observed heterogeneity.

DISCUSSION
To our knowledge, this is the first systematic review and metaanalysis to show that AI-based screening methods are highly sensitive and specific for the diagnosis of pathological myopia from fundus images.The area under the SROC curve was 0.9905, suggesting excellent classification performance of included models, as well as high diagnostic odds ratios (pooled DOR = 841.26)suggesting that included models generally possessed robust discriminative ability.However, the presence of unexplained statistical heterogeneity means that results should be interpreted with caution.
The majority of included studies used CNN-based models to detect PM.Three of the studies reported lower sensitivities [25,30,31], two of which used SVM to classify images, which would be expected to have lower discriminative ability than a deep learning-based approach.While one study employing SVM [37] was excluded from the quantitative analysis, it also showed a lower discriminative ability relative to CNN-based models.
Fagan nomograms were used to describe the likelihood of a patient having PM if the diagnostic tool deemed them to be a positive case (post-test probability); these demonstrated generally high (>85%) post-test probabilities, suggesting that diagnostic The potential application of AI models to diagnose PM in practice is multi-fold.Firstly, these analyses show a universally high observed diagnostic accuracy of AI-based models, reinforcing their capability as powerful screening tools for PM.
Secondly, our analyses reveal that the diagnostic accuracy and discriminative capability of these models is comparable to that provided by ophthalmologists, highlighting their capability as decision aids.In all seven studies comparing algorithm and grader performance, algorithm sensitivity and specificity was comparable with that of human graders, offering support for the use of such algorithms as screening or triage tools.The use of AI in an assistive capacity may reduce uncertainty in diagnosis and reduce the variability in the diagnosis made between healthcare professionals [41].
Tools based on AI models can work with clinicians to guide triage and referral decisions in general practice or non-specialist centres with a high case burden, as described by De Fauw et al. [42].This may have particular benefit for clinicians in training, or in regions with reduced incidence of PM, who may derive benefit from decision support in selecting cases of PM [43], or in areas with poor access to healthcare services [44].
Identification and close follow-up of patients with uncomplicated PM is crucial in enabling early management of treatable complications such as myopic CNV-for example with anti-VEGF therapy [45]-optimisation of visual acuity and stabilisation of progressive myopia.As novel treatments, such as stem cell therapy, gain prominence in the management of retinal disease [46], timely and targeted intervention is likely to be effective in reducing the public health burden of PM.Identification of PM can enable prognostication and careful multidisciplinary planning to mitigate the social, economic [47] and cognitive [48] impacts of progressive visual loss on the individual.
No included studies reported on the implementation of AIbased screening methods in clinical practice.While this review highlights the potential of AI to make highly specific and sensitive judgements on the presence or absence of pathological myopia, consideration must be given to generalisability across populations, explainability of screening decisions [49], and patient and healthcare professional acceptability [50].
Several existing reviews and meta-analyses examine the sensitivity and specificity of AI-based methods for the detection of other ophthalmic conditions from fundus photographs.Dong et al. [12] performed a systematic review and meta-analysis of AI algorithms used for the diagnosis of age-related macular degeneration, finding a pooled sensitivity and specificity of 0.88 and 0.90 respectively.Chaurasia et al. [13] demonstrated a pooled sensitivity and specificity of 0.92 and 0.94 respectively for the diagnosis of glaucoma from fundus images using AI algorithms.Finally, a metaanalysis by Wu et al. [51] observed a combined AUROC of 0.97-0.99 for the use of AI in diabetic retinopathy screening.
Examples are present in the literature of the use of AI for the detection of PM from optical coherence tomography (OCT) images [32,52].However, this was beyond the scope of this study.At the time of writing, OCT machines remain expensive, rendering them inaccessible in many regions.Fundus imaging, however, is widespread, and screening tools based on fundus photography may have a more significant clinical impact in less economically developed regions.
Future research assessing the diagnostic performance of AI models using OCT images for detection of PM may be useful in regions where the technology is widely used.

Strengths and weaknesses
The present study has several strengths.We used robust metaanalytic methodology to assess the pooled diagnostic accuracy of included studies, according to the PRISMA guidelines.Rigorous risk of bias assessment was performed, using two checklists, to identify studies which did not meet quality standards for inclusion, and heterogeneity and publication bias were comprehensively assessed using established methods.
Several limitations to this analysis are noted.First, there was significant statistical heterogeneity between the included studies, which was not entirely explained by analysis of threshold effects, study origin, training set size, study quality, the presence of external validation or the algorithm used.
A sensitivity analysis showed that studies deemed to have a moderate risk of bias reported a lower sensitivity and specificity, and studies without an external validation set reported a lower sensitivity-explaining some of the observed heterogeneity.Notably, there was considerable variation in the case-mix of positive to negative cases between studies, with PM prevalence varying from 3.78% in the test set used by Demir et al. to over 50% in other studies, potentially contributing to spectrum bias (where the discriminative ability of a diagnostic test varies according to the population in which it is used).
Meta-analyses of diagnostic accuracy of AI-based tools for the diagnosis of other ophthalmic conditions also demonstrate high unexplained heterogeneity [53], suggesting that variation in study and model design may have contributed to heterogeneity.However, it was not possible to assess the effects of variation in study design in detail (beyond subgroup analysis based on model type) due to the limited amount of methodological data provided in some studies.Regardless, the included studies spanned a diverse range of techniques and approaches, which-in the context of universally high diagnostic accuracy-suggests that AIbased techniques may possess excellent external validity for this purpose.
Secondly, several studies did not include an external validation set, limiting the generalisation ability of the algorithms reported.Ophthalmic patient populations are diverse, underlining the need for external validation, and meaning that the results of these studies should be interpreted with caution.
Third, there is a small degree of overlap between images contained within public datasets used in the studies, which could result in inflated estimates of diagnostic accuracy when comparing studies.Fourth, a retrospective study design was employed by all included studies.Retrospective design can lead to a selection bias, and have been shown to lead to overestimation of diagnostic accuracy [54].Fifth, there was a preponderance of studies from East Asian countries, particularly China, in this review and meta-analysis.Planned subgroup analysis assessing studies by country of origin showed a small increase in sensitivity and specificity for studies originating from China, however it is unknown whether this is due to model factors or ethnic differences in fundus appearance [55].
Finally, due to the limitations of the available data, we did not include an analysis of model performance by severity of pathological myopia.It is therefore unknown how disease severity affects diagnostic accuracy in this case.Evidence suggests that use of severe cases may result in somewhat inflated estimates of diagnostic accuracy, and this would be an appropriate area for future research [54].

CONCLUSION
This systematic review and meta-analysis provide robust early evidence for the diagnostic accuracy of AI-based tools in the diagnosis of PM.Such tools are likely to have significant impact in screening, triage, assisted diagnosis and monitoring of myopic patients, and may enable earlier diagnosis and improve clinical outcomes for patients at risk of developing PM.

AUROC
area under receiver-operator curve, PPV positive predictive value, NPV negative predictive value, LR-negative likelihood ratio, LR+ positive likelihood ratio, F1 F1-Score, CNN convolutional neural network, SVM support vector machine.a

Fig. 1
Fig. 1 PRISMA flowchart showing study design.Added numbers (denoted with '+') represent studies added in the second database search.

Fig. 3
Fig. 3 Forest plot and SROC curves for main analysis.Forest plot of sensitivities (a) and specificities (b) and SROC curve for overall diagnostic performance (c) for all studies included in meta-analysis.CI indicates confidence interval; SROC, Summary receiver operating characteristic, AUC indicates area under the curve, SE indicates standard error; Q* indicates where sensitivity = specificity (intersection of diagonal with SROC curve).

Table 1 .
Summary of included studies.

Table 1 .
continued NSNS not stated.a Studies excluded from the meta-analysis as no contingency table could be formed.b From 4432 eyes.

Table 2 .
Model type and pooled performance for included studies.