Image-Based Deep Learning for Cataract Diagnosis: Systematic Review and Meta-Analysis

Introduction

The lens is a biconvex, transparent structure in the anterior segment of the eye, which focuses light to project images of objects at varying distances onto the retina. A cataract, defined as clouding of the lens, is primarily an age-related degenerative disease, and congenital and pediatric cataracts also occur [1]. Early cataracts are asymptomatic, but progressive clouding can lead to visual impairment, greatly reducing quality of life and productivity [2]. Cataracts are the major cause of visual loss worldwide. In 2020, the estimated number of the blind aged 50 years and older was 15 million, and moderate to severe visual disorders due to cataracts affected an estimated 79 million people aged 50 years and older [2]. These figures suggest the number of patients with blindness and moderate to severe visual disorders rose by 30% and 93%, respectively, versus the year 2000 [3]. Blinding cataracts refer to cataracts causing severe visual impairment (visual acuity ≤3/60) or blindness (visual acuity <3/60) according to the WHO (World Health Organization) and ICD-11 (International Classification of Diseases, 11th Revision) [4]. About 94% of blinding cataracts occur in low- and middle-income countries (LMICs), and cataract-related visual disorders are causally related to poverty in low-resource settings (LRS) [1]. As the conventional diagnosis of cataracts relies on ophthalmologists and complex equipment, it is subjective and requires health care resources, making it difficult to satisfy the need for large-scale screening. Moreover, due to the limitations of infrastructure and the lack of trained personnel in LMICs, inequalities are present in early cataract detection and severity classification.

Deep learning (DL) can process pixel-level information that cannot be recognized by the human eye to greatly contribute to the analysis of medical images, assist doctors in clinical decision-making, and enhance screening efficiency. As a branch of artificial intelligence (AI), DL models are inspired by the brain and specialized in pattern recognition [5]. Therefore, DL has been rapidly evolving with broad prospects in the field of medical image analysis involving disease detection, classification, segmentation, and image registration [6]. For example, convolutional neural networks (CNNs), as the primary technique of DL for image learning, perform excellently in image classification and feature extraction, making it a cornerstone in medical imaging [7]. Residual network (ResNet), one of the landmark architectures of CNNs, has contributed to the development of DL, especially with outstanding performance in benchmarking in image recognition and classification [8]. The superiority of the ResNet model in the medical field lies in its capability to efficiently train the deep network and raise the accuracy of image recognition [9].

The strategic importance of AI is to raise the quality of care and potentially reduce costs in high-income economies and to address critical health care issues and staffing shortages and provide access to specialized skills in LMICs [10]. Nowadays, DL is developed mostly based on data from high-income countries and regions and relies on high-resolution images and advanced electronic devices. In remote regions, however, fragmented health care systems are generally characterized by insufficient infrastructure, a shortage of professionals, and a lack of health care resources, so it is difficult to guarantee the quality of screening and diagnosis. The following problems are present in the available DL models: (1) lack of compatibility: mismatch with the low-cost equipment used in LRS [10]; (2) lack of generalization capability: significant decline in the model performance in real-world scenarios, especially under different lighting conditions or in different patient populations [11]; (3) lack of clinical validation: a recent systematic review of studies assessing the use of AI algorithms for medical image analysis found that only 6% of the included studies (n=516) conducted external validation, and the assessment of the diagnostic efficiency was lacking [12]; and (4) training data bias and lack of diversity in data (race, age, and disease subtypes) weaken the model’s generalization capability [13].

In the past 5 years, breakthroughs have been made in the use of DL for ophthalmic image analysis in diabetic retinopathy (DR), retinopathy of prematurity, and glaucoma. By automating the massive processing of ophthalmic images, DL can achieve a more accurate and rapid diagnosis of cataracts, reducing the doctors’ subjectivity and errors of conventional methods. Moreover, DL can fuse multiple image data (eg, slit lamp, fundus, and optical coherence tomography [OCT] images) to make a comprehensive and accurate diagnosis through multimodal image analysis. In addition, diagnostic results can be obtained quickly from the DL model with real-time diagnostic capability, greatly improving work efficiency and achieving large-scale screening and early diagnosis, especially in primary health care institutions and LRS [14]. Many DL-based diagnostic tools have been approved by the US Food and Drug Administration (FDA), but they require further evaluation and independent quality review [15]. For example, IDx-DR [16,17], the first AI system approved by the FDA, is a CNN-based system for automated screening for DR, with high sensitivity and specificity, which can contribute to the early diagnosis and lower the risk of visual loss in clinical practice, especially in LRS.

Image-based DL has exhibited greater potential for automatic cataract detection and classification using fundus and slit lamp images. However, the existing findings are still heterogeneous, and a systematic review of DL algorithms for cataract image analysis is still lacking. Therefore, this study was conducted to systematically assess the performance of different DL models in cataract detection and classification from sensitivity, specificity, and the area under the ROC curve (AUC), thereby revealing the methodological and reporting quality and contributing to the clinical translation of DL.

Methods

Registration and Study Design

This study adhered to the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-analyses of Diagnostic Test Accuracy Studies) [18], and the study protocol was registered on PROSPERO (International Prospective Register of Systematic Reviews; CRD420251030230). We acknowledge that the registration was completed on April 10, 2025, following the initial literature search conducted on April 1, 2025; thus, this is a retrospective registration.

Search Strategy, Eligibility Criteria, and Data Extraction

Based on the predefined criteria, 2 investigators (RXL and HYL) independently searched Web of Science, IEEE Xplore, Embase, PubMed, and Cochrane Library up to April 1, 2025, for studies published from 2019 to 2025.

The retrieved studies were first imported into EndNote to eliminate duplicate publications. Then the title and abstract were read, and the full text of clearly or potentially eligible studies was examined. No restrictions were imposed on the geographical location or study setting. The following study types were excluded, including letters, non–peer-reviewed reports, narrative reviews, animal studies, and conference abstracts. The search strategy is shown in Multimedia Appendix 1, with specific search phrases, Boolean operations, and field restrictions. The following studies were excluded during full-text screening: studies failing to report key values such as true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN), making it impossible to make a contingency table; studies reporting only composite metrics like sensitivity, specificity, or AUC without providing the underlying data; and studies that reported data ambiguously or in a manner that prevented accurate construction of a 2×2 contingency table. The following is a table of inclusion and exclusion criteria (Table 1). Discrepancies were resolved through discussion with a third investigator (SL or LHL) if needed.

Only studies on the performance of image-based DL algorithms in cataract detection and classification were included following the eligibility criteria (Table 2).

Two investigators (RXL and HYL) independently extracted the following data using standardized data extraction tables: basic characteristics (country, publication year, study site, and type), dataset characteristics (nature, number of images, and presence or absence of external validation), and performance metrics (sensitivity and specificity). Discrepancies were settled by discussion with a third investigator. Contingency tables were used to directly extract the data on binary diagnostic accuracy, including TP, FP, TN, and FN. During data extraction, double-checking was performed, and the original authors were contacted to obtain supplementary data. These data were then used to calculate pooled sensitivity, specificity, and other metrics. If a study provided multiple contingency tables for the same or different DL algorithms, they were assumed to be independent of each other.

Primary and Secondary Outcomes

Primary and secondary outcomes were defined to assess the performance of DL in cataract detection and classification. The primary outcomes included sensitivity, specificity, and positive and negative likelihood ratios.

The secondary outcomes were used to assess the accuracy of DL versus machine learning (ML) algorithms in cataract diagnosis through subgroup analyses and compare DL algorithms with human experts in studies using identical datasets. The datasets in each study were also investigated for the reference standard used to determine whether transfer learning was applied, the methods for model testing and validation, and the sources and characteristics of the datasets (Table 3).

Quality Assessment

Three investigators independently assessed the risk of bias (RoB) and applicability concerns in the clinical context in the included studies by using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) [82]. Cross-validation was implemented to enhance interrater agreement, and the methodology for consistency assessment was explicitly reported. Deeks’ funnel plots were drawn to evaluate the publication bias if more than 10 studies were included. Statistical significance was set at P<.05. Deeks’ funnel plot asymmetry test was performed using the Deeks command within the MIDAS package in STATA 18.0 (StataCorp LLC).

Statistical Analysis

STATA 18.0 and RevMan 5.4 (Review Manager; The Cochrane Collaboration) were used for data analyses. The summary receiver operating characteristic (SROC) curve plotting, assessment of heterogeneity, and analysis of publication bias were performed using STATA 18.0 to enhance transparency and reproducibility. The heterogeneity was assessed by the Cochran Q test and I² statistic. The threshold of the I² statistic for quantifying heterogeneity proposed by Higgins et al [83] was adopted: I²≤25%: low, I²≈50%: moderate, I²≥75%: substantial. This strategy was specifically designed to identify whether inconsistencies in study results stemmed from random factors or reflected substantive discrepancies. When significant heterogeneity (P<.05 or I²>50%) was identified [84], a bivariate mixed effects model was adopted. The bivariate random-effects model was fitted using the MIDAS package in STATA 18.0. This method demonstrated particular efficacy in synthesizing pooled estimates of sensitivity, specificity, and AUC across studies. Its advantage resides in the capability to systematically address metric variability while preserving the intrinsic correlation.

To evaluate the accuracy of the DL algorithm, a hierarchical SROC curve was fitted. We calculated corresponding 95% CIs for sensitivity, specificity, and AUC using the Delta method, which linearized the nonlinear relation of the log-transformed sensitivity and specificity by a first-order Taylor expansion. Then the variance-covariance matrix of the parameter estimates was propagated, whereas prediction intervals (PIs) incorporated between-study heterogeneity by modeling the covariance structure of sensitivity and specificity [84,85]. Sensitivity analyses were conducted on all DL algorithms to be evaluated, rather than only the one with the highest accuracy. A random-effects model was used to explain potential between-study variability.

Ethical Considerations

This study required no informed consent or ethical approval. Data previously collected from human subjects in ethical/institutional review board–approved studies were used. All studies included adhered to the Declaration of Helsinki.

Results

Search Results

Initially, 2235 studies were retrieved, of which 492 duplicates were excluded. Following study screening, 1680 studies were excluded from quantitative synthesis (meta-analysis), including 1617 studies involving animal research, nondisease studies, surgical technique investigations, reviews, and conference reports; 10 studies not using deep learning algorithms; 46 studies lacking sufficient data for constructing 2×2 contingency tables or reporting data in formats incompatible with pooling (eg, AUC only); 4 studies not focusing on cataract diagnostic models; and 3 studies that did not address cataract disease (Figure 1; Checklist 1).

‎

Study Characteristics

As shown in Tables 2-4, all the included studies were published in 2019‐2024, involving 171,416 images. Retrospective data were used in 46 studies, prospective data in only 5 studies, and cross-sectional data in 12 studies. The data came from open-access sources in 31 studies. The sample size was prespecified in one study, and low-quality images were excluded in 15 studies. External validation was conducted in 12 studies, while the remainder performed internal validation. Six studies compared DL with traditional ML models using the same dataset, while another 6 compared DL models with human experts. Forty-four studies focused on cataract detection, 17 on cataract classification, and 2 on both detection and classification. Cataract detection was categorized as binary detection (presence vs absence of cataracts, n=21) and multidisease detection (n=25). Moreover, cataracts were classified into mild (n=18), moderate (n=11), and severe (n=19) types. Additionally, 49 studies did not describe cataract classification; among the other studies, clinical subtypes included posterior subcapsular cataract (PSC; n=4), pediatric cataract (n=2), posterior polar cataract (PPC; n=1), nuclear cataract (NC; n=10), and cortical cataract (CC; n=6).

Pooled Performance of DL Algorithms

Finally, 63 studies [19-81] with sufficient data (97 contingency tables) were included for the assessment of DL performance in cataract detection and classification [86]. Hierarchical SROC curves for cataract detection (45 contingency tables) and classification (52 contingency tables) are provided in Figures 2A and 3A, respectively. The classification task involved multiclassification (eg, mild, moderate, and severe cataracts), and separate SROC curves were generated for each category. For cataract detection, the pooled sensitivity and specificity of DL were 96% (0.95‐0.97) and 98% (0.97‐0.99), respectively, with an AUC of 0.99 (0.98‐1.00). For cataract classification, DL had pooled sensitivity and specificity of 94% (0.93‐0.96) and 97% (0.96‐0.98), respectively, with an AUC of 0.99 (0.98‐1.00). Great heterogeneity and inconsistency were observed across cataract severity (mild: I²=99%, moderate: I²=96%, severe: I²=99%; P<.001), suggesting substantial variability in diagnostic or methodological methods across studies. 25 contingency tables for mild cataracts were used in 18 studies, 16 contingency tables for moderate cataracts in 11 studies, and 28 contingency tables for severe cataracts in 19 studies, which may introduce classification imbalance, potentially influencing the performance assessment.

‎

‎

One or more DL algorithms were reported in most studies, and DL with the highest accuracy was selected, ultimately obtaining 53 contingency tables. For cataract detection, DL had pooled sensitivity and specificity of 97% (96%‐98%) and 98% (97%‐99%), respectively, with an AUC of 0.99 (0.98‐1.00) (Figure 2B). For cataract classification, the pooled sensitivity and specificity were 95% (0.92‐0.97) and 98% (0.96‐0.99), respectively, with an AUC of 0.99 (0.98‐1.00) (Figure 3B). Threshold analyses were conducted using STATA 18.0 to investigate threshold effects. For diagnostic models in primary studies that did not prespecify a threshold, sensitivity and specificity corresponding to the optimal threshold reported in the study were extracted. If a primary study reported results for multiple thresholds, we prioritized extracting and analyzing data corresponding to the threshold associated with the reported primary endpoint or optimal operating point. The spatial distribution of classification thresholds in ROC curves was analyzed; each threshold corresponded to a unique point on the ROC curve, and systematic evaluation of these points could help us quantify performance variability across thresholds and detect threshold-driven instability (defined as significant performance fluctuations within narrow threshold ranges). The SROC curve displayed no “shoulder-arm” distribution (Figures 2 and 3).

However, the model’s performance in independent external datasets (detection: sensitivity 87%, specificity 93%; classification: sensitivity 89%, specificity 90%) was lower than the overall estimates. Notably, the lower performance observed in external validation datasets compromised the generalization capability of the model, potentially attributable to domain shift, warranting caution when applied to new populations or settings.

Subgroup Analyses

Overview

Traditional ML or DL algorithms were reported in the included studies. These studies varied in primary objectives (detection, classification, or both detection and classification). Due to overlapping objectives among studies, the sum of the number of studies on traditional ML and DL algorithms did not match the total number of included studies. According to the Lens Opacities Classification System III (LOCS III) [87] and the methods of Mackenbrock et al [88] and Gali et al [89], we classified cataracts into mild, moderate, and severe types and further categorized cataracts into PSC, pediatric cataract, PPC, NC, and CC.

Detection

Algorithm Types

DL algorithms were described in 36 studies (45 contingency tables). The pooled sensitivity and specificity of DL were 96% (95%‐97%) and 98% (97%‐99%), respectively, with an AUC of 0.99 (0.98‐1.00; Figure 2A). Additionally, traditional ML algorithms were described in 5 studies (13 contingency tables). The pooled sensitivity and specificity of ML were 90% (87%‐91%) and 94% (91%‐96%), respectively, with an AUC of 0.95 (0.93‐0.97; Figure 4A).

‎

Disease Types

Around 37 studies (57 contingency tables) focused on nonspecific or general cataracts; the pooled sensitivity, specificity, and AUC were 96% (94%‐97%), 98% (97%-98%), and 0.99 (0.98‐1.00). The Cochran Q test revealed a statistically significant result (Q=199.023; P<.001) and high inconsistency (I²=99%; 95% CI 98‐99), indicating significant between-study variability. For cataract detection (CC: one study with one contingency table, pediatric cataract: 2 studies with eight contingency tables, NC: 2 studies with 3 contingency tables, PSC: one study with one contingency table, PPC: one study with one contingency table), the pooled sensitivity and specificity were 93% (90%‐95%) and 96% (92%‐98%), respectively, and the AUC was 0.97 (0.95‐0.98) in 4 studies (13 contingency tables). Due to the extremely limited sample sizes across clinical subtypes, with most subtypes reported by only one study, the meta-analysis results were less robust. Consequently, only a pooled heterogeneity assessment could be conducted. The Cochran Q test revealed a statistically significant result (Q=9.167; P=.005) and high inconsistency (I²=78%; 95% CI 53‐100), indicating significant between-study variability (Figures 5A and 5B).

‎

Validation Types

Great heterogeneity and inconsistency were observed in the validation methods for cataract detection (I²=99% for both internal and external validation; P<.001), suggesting substantial variability in diagnostic or methodological methods across studies. The Deeks’ funnel plots revealed no great publication bias in the internal validation (P=.68) or the external validation (P=.23). The internal validation was conducted in 38 studies (54 contingency tables), which showed a pooled sensitivity of 96% (95%‐97%) and a pooled specificity of 98% (97%-98%), with an AUC of 0.99 (0.98‐1.00). The external validation was performed in only 8 studies (15 contingency tables), which showed a pooled sensitivity of 87% (81%‐92%) and a pooled specificity of 93% (86%‐96%), respectively, with an AUC of 0.95 (0.93‐0.97; Figures 6A and 6B).

‎

Classification

Algorithm Types

DL algorithms were described in 17 studies (52 contingency tables); the pooled sensitivity and specificity of DL were 94% (93%‐96%) and 97% (96%‐98%), respectively, with an AUC of 0.99 (0.98‐1.00; Figure 3A). Additionally, traditional ML algorithms were described in 6 studies (15 contingency tables); the pooled sensitivity and specificity of ML were 88% (85%‐90%) and 94% (90%‐96%), respectively, with an AUC of 0.94 (0.92‐0.96; Figure 4B).

Disease Types

For mild cataracts, 18 studies (25 contingency tables) had a pooled sensitivity of 92% (89‐94%) and a pooled specificity of 96% (94%‐97%), with an AUC of 0.98 (0.96‐0.99). For moderate cataracts, 11 studies (16 contingency tables) found comparable performance: SE=94% (90%‐96%), specificity=97% (95%‐98%), and AUC=0.99 (0.97‐0.99). For severe cataracts, 19 studies (28 contingency tables) had a sensitivity of 93% (90%‐95%) and a specificity of 98% (96%‐99%), with an AUC of 0.98 (0.97‐0.99). Great heterogeneity and inconsistency were observed across cataract severity (mild: I²=99%, moderate: I²=96%, severe: I²=99%; P<.001), suggesting substantial variability in diagnostic or methodological methods across studies. Furthermore, eight studies with 28 contingency tables focused on nonspecific or general cataracts, which had a pooled sensitivity of 95% (92%‐96%) and a pooled specificity of 98% (97%‐99%), with an AUC of 0.99 (0.98‐1.00). Great heterogeneity and inconsistency were observed across cataract severity (nonspecific or general cataract: I²=96%, NC: 100%; P<.001), suggesting substantial variability in diagnostic or methodological methods across studies. In the NC subgroup (9 studies with 27 contingency tables), extremely high heterogeneity was observed (I²=100%). Consequently, to maintain statistical consistency, we did not calculate a pooled estimate for this subgroup. Instead, the descriptive analysis showed a sensitivity of 89%‐93% and a specificity of 93%‐97% across these studies. Four studies with 15 contingency tables considered other clinical subtypes (PSC: one study with 3 contingency tables; CC: 4 studies with 12 contingency tables), with pooled sensitivity, specificity, and AUC of 91% (84%‐95%), 96% (94%‐97%), and 0.98 (0.96‐0.99). Due to the extremely limited sample sizes of other clinical subtypes, the meta-analysis results were less robust. Consequently, only a pooled heterogeneity assessment could be conducted. The Cochran Q test revealed a statistically significant result (Q=120.355; P<.001) and high inconsistency (I²=98%; 95% CI 97‐99), indicating significant between-study variability (Figures 7A-7E).

‎

Validation Types

The internal validation was adopted in 17 studies (59 contingency tables), which showed a pooled sensitivity of 93% (92%-95%) and a pooled specificity of 97% (96%‐98%), with an AUC of 0.99 (0.97‐0.99). The external validation was adopted in only 4 studies (9 contingency tables), which showed a pooled sensitivity of 89% (83%‐92%) and a pooled specificity of 90% (86%‐92%), with an AUC of 0.95 (0.93‐0.97; Figure 8).

‎

Comparison Between DL and ML Algorithms

Detection

Two studies compared DL (8 contingency tables) and ML algorithms (7 contingency tables) using the same dataset. The pooled sensitivity was 94% (92%‐96%) for DL and 91% (85%‐95%) for ML algorithms. The pooled specificity was 99% (97%‐100%) for DL and 90% (86%‐93%) for ML algorithms. The AUC was 0.97 (0.95‐0.98) for DL and 0.96 (0.94‐0.97) for ML algorithms. For DL algorithms, the Cochran Q test revealed a statistically significant result (Q=9.675; P=.004) and high inconsistency (I²=79%; 95% CI 55‐100). For ML algorithms, the Cochran Q test also revealed a statistically significant result (Q=5.853; P=0.03) and high inconsistency (I²=66%; 95% CI 23‐100). Therefore, both DL and ML exhibited significant heterogeneity. However, due to overlapping CIs and no direct between-group comparisons, the statistical significance of the between-group heterogeneity could not be conclusively established based on these findings. The comparison between DL and ML algorithms for cataract detection was constrained at the time of this writing by minimal validation evidence, with only 2 studies available for direct benchmarking. As a result, it severely limited statistical power, compromised generalization capability, and inflated performance estimates for DL models (Figures 9A and 9B).

‎

Classification

Four studies compared DL (10 contingency tables) and ML algorithms (10 contingency tables) using the same dataset. The pooled sensitivity was 95% (91%‐97%) for DL and 89% (86%‐91%) for ML algorithms. The pooled specificity was 98% (95%‐99%) for DL and 94% (89%‐97%) for ML algorithms. The AUC was 0.99 (0.98‐1.00) for DL and 0.94 (0.92‐0.96) for ML algorithms (Figures 9C and 9D).

DL Algorithms Versus Human Experts

Direct comparisons between DL algorithms (7 contingency tables) and human experts (10 contingency tables) were performed across 7 studies using the same datasets. For DL algorithms, quantitative pooling was not conducted due to extreme heterogeneity (Q=933.852; P<.001; I²=100%). Instead, a descriptive analysis revealed highly variable performance of DL algorithms, with sensitivity estimates ranging from 72% to 93% and specificity ranging from 64% to 99%. In contrast, human experts demonstrated moderate heterogeneity (Q=5.811; P=0.03; I²=66%), allowing for a pooled analysis. The pooled sensitivity and pooled specificity for human experts were 93% (95% CI 77%‐98%) and 95% (95% CI 79%‐99%), respectively, with an AUC of 0.98 (95% CI 0.97‐0.99; Figure 10). These results highlight that DL algorithms exhibited significantly higher heterogeneity than human experts. The Deeks’ funnel plot indicated no significant publication bias for human experts (P=.16), but potential borderline publication bias for DL algorithms (P=.05). The extreme heterogeneity (I²=100%) observed in the DL algorithms underscores that their efficacy is highly dependent on specific study conditions and architectures, precluding a uniform performance metric.

‎

Publication Bias and Heterogeneity

The Deeks’ funnel plot revealed no significant publication bias for cataract detection (P=.48). However, potential borderline publication bias was detected for cataract classification (P=.05). However, the wide distribution of studies near the regression line in the plot should be further considered (Figures 11A and 11B).

‎

The included studies showed substantial heterogeneity. The high I² values (sensitivity: 95.00% and specificity: 97.11% for cataract detection; sensitivity: 95.94% and specificity: 98.55% for classification) suggested considerable between-study variability (P<.001), which was further explored by subgroup analyses (Figures 12 and 13).

‎

‎

RoB

The quality of included studies was assessed using QUADAS-2, and a summary of findings is displayed in Figure S1 in Multimedia Appendix 1 (summary of methodological quality of 63 studies included). A detailed assessment for each item of RoB and applicability concern is also provided in Figure S2 in Multimedia Appendix 1 (diagrams of methodological quality of 63 studies included). High RoB or selection bias was detected in 34 studies due to a lack of randomization or eligibility criteria in the patient selection domain; RoB was high or unclear in 44 studies due to no predefined threshold in the index test domain.

Due to inconsistencies in reference standard (no reporting of whether blinding was implemented and the presence or absence of a predefined threshold), RoB was high or unclear in 2 studies. RoB was high or unclear in 34 studies in the domain of flow and timing due to no mention of whether the same gold standard was used or the presence or absence of an appropriate time gap.

High or unclear applicability in the domain of patient selection was detected in one study and unclear applicability in the domain of index test in 5 studies; 2 studies showed applicability concerns in the domain of reference standard.

Discussion

This meta-analysis included 63 studies and rigorously assessed the study quality using the QUADAS-2 to summarize the evidence to date on the performance of image-based DL in cataract diagnosis. The results revealed that DL algorithms might offer higher accuracy than traditional ML algorithms and fall within the range of reported accuracy of human experts in the detection of cataracts, demonstrating potential as tools for automated diagnosis. However, given the moderate quality and high heterogeneity of the current evidence base, these DL algorithms are considered primarily as adjuncts to cataract diagnosis.

DL has made tremendous progress in automated image analysis [6]. In clinical practice, a severe imbalance between supply and demand is present in ophthalmologic diagnosis. As AI advances, DL is expected to raise diagnostic efficiency and thus help alleviate health care resource inequality.

Four relevant systematic reviews and meta-analyses were identified. (1) Cheung et al [90] found high sensitivity, specificity, and reproducibility and great heterogeneity of the ML model for cataract diagnosis in children and adults, but the strength of evidence was limited since only 11 studies with 13 contingency tables were included. (2) Liu et al [91] found comparable performance of DL and experts in medical imaging, but only 2 out of 18 ophthalmology studies involved cataracts, necessitating in-depth research on DL in cataract diagnosis. (3) Aggarwal et al [92] included 82 ophthalmology studies, which did not involve cataracts, and verified that DL possesses good sensitivity, specificity, and AUC for the feature identification of other eye diseases, but the algorithmic evaluation criteria remain to be standardized due to high heterogeneity. (4) Islam et al [93] demonstrated that DL can be applied to retinal vessel segmentation, so it can be popularized in LMICs. Therefore, methodological limitations should be critically assessed during the clinical translation of DL to further improve its reliability.

This meta-analysis systematically evaluated the effectiveness of DL versus traditional ML in cataract detection and classification to provide a basis for clinical decision-making. For cataract detection, DL had a pooled sensitivity of 96% (95% CI 95%‐97%) and a pooled specificity of 98% (97%‐99%), with an AUC of 0.99 (0.98‐1.00); traditional ML had a pooled sensitivity of 90% (87%‐91%) and a pooled specificity of 94% (91%‐96%), with an AUC of 0.95 (0.93‐0.97). For cataract classification, the pooled sensitivity and specificity of DL were 94% (93%-96%) and 97% (96%‐98%), with an AUC of 0.99 (0.98‐1.00); the pooled sensitivity and specificity of traditional ML were 88% (85%‐90%) and 94% (90%‐96%), with an AUC of 0.94 (0.92‐0.96). Available evidence suggests that DL models exhibit high sensitivity and specificity in automated cataract diagnosis and that medical image-based DL still demonstrates superior robustness to ML despite the RoB in most studies. However, our conclusions were made partly based on studies of low quality due to a lack of external validation and nonstandardized reporting of performance metrics, which may overestimate algorithmic accuracy. The use of overlapping public repositories introduced potential clustering bias. To ensure robust generalizability, future research must prioritize validation on independent, multicenter, and nonpublic datasets. Furthermore, the lack of a sensitivity analysis excluding high-RoB studies implied that the high overall performance of DL models could be partially influenced by methodologically weaker studies, warranting validation in future high-quality trials. Finally, due to the limited number of eligible studies, we pooled smartphone-based data from varying modalities, including diffuse photography and slit-lamp adapters. While this aligned with our assessment of general mobile health accessibility, it introduced potential optical heterogeneity. Consequently, statistical quantification of performance differences between optical sectioning and diffuse lighting was not feasible, warranting future separate investigations.

We identified the potential of DL for clinical use from the included studies, but the number of studies directly comparing the performance of DL and ML in cataract diagnosis was limited, with only 2 in detection and 4 in classification. Traditional ML (eg, logistic regression, random forest, and support vector machine) is dependent on manual feature engineering and classifiers, whose performance is limited by feature quality and domain knowledge [94]. By contrast, DL (eg, CNN and transformer-based vision models) has higher accuracy in image recognition by automatic feature extraction but requires higher data volume and stronger arithmetic support [95].

The number of included studies was limited, so image segmentation performance was not meta-analyzed. Image segmentation is essentially a pixel-level prediction task, which relies on the ability to characterize discriminative features. However, it is often difficult for traditional ML to effectively capture such features. In contrast, DL demonstrates significantly better performance by automated learning of complex patterns in highly heterogeneous data [6]. The property of DL models is such that they contain millions of parameters that need to be optimized by data-driven training, while traditional ML models display higher stability in small-sample multiclassification [95]. These findings suggest that traditional ML models may have the potential to raise the accuracy in cataract classification, while DL is more advantageous in reducing misdiagnosis rates. A combination of the 2 may facilitate the clinical translation of AI in the future.

Despite the available evidence on the clinical translational potential of DL, only 7 studies compared the efficacy of DL and human experts in cataract detection, and no studies compared their performance in cataract classification. Notably, significant interrater variability was present in expert performance due to heterogeneity in the cumulative clinical experience and health care resource allocation, highlighting the need for more comprehensive comparative studies. These studies generally report positive results of DL, but an optimism bias in algorithm performance may be produced due to insufficient sample sizes and methodological heterogeneity. Therefore, it is urgently needed to adopt a standardized study design and transparent reporting norms for improving the quality of evidence and clinical translational value of DL. DL has been successfully applied to retinal fundus image analysis [14] for automated diagnosis of DR [96] and glaucoma [97]. Considering the technical feasibility of DL in ophthalmic disease screening, it is recommended that DL serve as an adjunct to clinical diagnosis to optimize the treatment process by human-machine collaboration. Additionally, no great publication bias was detected, but the findings need to be interpreted with caution. It is recommended that a clinician’s diagnostic efficiency be used as a core assessment metric and that a real-world validation be incorporated in subsequent studies.

In the field of AI methodology, several standardized guidelines have been recently issued [98,99]. However, no unified consensus has been reached on AI for cataract diagnosis. At the time of this writing, computer-aided diagnostic techniques mainly use a combination of medical image processing and AI, but research focuses mostly on eye diseases of the posterior segment (eg, DR, and age-related macular degeneration) and less on the anterior segment (eg, cataracts). Nowadays, cataract diagnosis relies on slit-lamp microscopy of lens morphology, and cataract classification is based on the LOCS III [87], the Oxford Clinical Cataract Classification and Grading System [100], and the American Cooperative Cataract Research Group method [101]. DL has made a preliminary breakthrough in automated cataract classification, such as the turbidity-density-location assessment system developed by Lin et al [19] based on anterior segment slit lamp images and the DL classification model developed by Zhou et al [102] using retinal fundus images, but a standardized severity scale adapted to DL is urgently needed for its clinical use. Therefore, it is recommended that future studies standardize the criteria for performance assessment of DL algorithms, with a focus on improving the transparency of methodology reporting and repeatability validation process.

Data scarcity and lack of generalization capability are key scientific challenges for DL. In this systematic review, retrospective designs were adopted in most of the included studies, while prospective designs and multicenter clinical trials accounted for less than 5%, and their annotation criteria were not optimized for the need of DL, restricting methodological rigor. We should recognize that most of the studies used double-blind, annotated, and quality-controlled image data for model training, which effectively improved diagnostic accuracy and reduced the RoB while balancing data size and quality. The included studies generally used data enhancement techniques, indicating the lack of high-quality annotated datasets and prospective validation studies.

The acquisition of representative data for clinical validation remains the major bottleneck. The model’s clinical generalization capability is severely restricted due to the long time consumption of pixel-level annotation of fundus images, domain shift resulting from cross-device and cross-race variations, and a lack of fine-grained pathology characterization. In the future, research should focus on multimodal data fusion for modeling (eg, OCT plus fundus images), deployment of edge computing architectures, and embedding of causal inference modules, thereby establishing intelligent diagnosis and treatment systems with both clinical credibility and engineering practicality.

In this study, great methodological heterogeneity was found in the reporting of DL performance metrics, including incomplete reporting of sensitivity or specificity, a lack of 2×2 contingency tables, and general overreliance on aggregated metrics such as AUC-ROC and F₁-score. Of particular note is that high AUC values (>0.90) may mask the risk of misjudgment of key positive events (eg, progressive cataract) in clinical scenarios with a severe imbalance in category distribution. Based on these findings, it is recommended that the confusion matrix serve as a core reporting metric. The above-mentioned problems can be gradually settled by high-quality studies.

Another key obstacle is no consensus on the interpretability of DL decision-making mechanisms. It is difficult to intuitively understand the decision-making logic of DL models due to their complex network structure and nonlinear feature extraction, which, as a “black-box” characteristic, has been deemed a challenge for clinical use. Recently, researchers have gradually revealed the intrinsic mechanism of DL models by gradient-weighted class activation mapping, adversarial testing, and causal inference [95]. For example, Chang et al [103] used adversarial samples to analyze the decision-making basis of DL models in glaucoma detection. Araújo et al [104] located the key region of DR in fundus images by multiple-instance learning. Explainable AI (XAI) is breaking through the limitations of traditional AI by synchronizing decision results with attributional explanations [105]. Abràmoff et al [106] systematically reviewed the interpretability framework for DL models in the medical field, laying a theoretical foundation for their clinical translation. Future research needs to further explore the innovative methods of XAI in medical image analysis to enhance clinical credibility.

This study showed that DL was applied primarily to the screening of eye diseases such as DR for which mature diagnostic guidelines have been established. Notably, differences in health care resource allocation should be considered when popularizing DL. In subsequent studies, the clinical effect and health economic benefits should be assessed across different DL algorithms, and the “black-box” problem of DL algorithms should be solved using interpretability methods to enhance clinical acceptance.