Age-related macular degeneration (AMD) remains a leading cause of irreversible blindness in older individuals globally [1]. Clinically, the disease is classified into dry AMD (dAMD), characterized by the progressive accumulation of drusen and geographic atrophy, and wet AMD (wAMD), which involves rapid vision loss due to macular neovascularization [2]. As the global population ages, the prevalence of AMD is projected to rise significantly; recent estimates indicate that the number of individuals with AMD-related vision impairment will increase from 8.06 million in 2021 to approximately 21.34 million by 2050 [1]. Consequently, early and accurate diagnosis is paramount. Timely detection allows for appropriate intervention, which is critical for slowing disease progression, preserving visual function, and improving overall patient prognosis.

Conventionally, color fundus photography (CFP) and optical coherence tomography (OCT) serve as the cornerstones for AMD screening and diagnosis. However, reliance on these modalities presents distinct challenges. CFP is frequently limited by image quality; issues such as media opacities or small pupils can render images ungradable, with rates as high as 47.6% in some screening contexts, and CFP often lacks sensitivity for detecting subtle early-stage structural changes or neovascular activity [3,4]. Conversely, while OCT offers high gradability (up to 97.7%) and detailed cross-sectional visualization, it is constrained by a limited field of view and reduced efficacy in identifying pigmentary abnormalities compared to CFP [3,4]. Beyond these technical constraints, the manual interpretation of vast imaging datasets is inherently labor-intensive, subjective, and prone to interobserver variability, creating a scalability bottleneck for population-wide screening.

In response to these challenges, deep learning (DL) algorithms using OCT, CFP, or multimodal imaging have emerged as a transformative approach, offering the potential for automated, high-throughput classification [5,6]. While these algorithms demonstrate theoretical superiority in efficiency and feature extraction, the current literature reveals substantial heterogeneity in performance outcomes [7,8]. Discrepancies regarding model generalization to real-world settings and the comparative performance of DL algorithms against ophthalmologists remain unresolved [7,8]. Two pivotal questions persist: First, how does the diagnostic performance of DL models quantitatively compare against ophthalmologists of varying expertise? Second, what factors, such as imaging modality, type of validation, database source, study centers, and unit of analysis, influence DL performance? Existing literature offers fragmented and sometimes contradictory insights, lacking a comprehensive quantitative synthesis.

Several previous meta-analyses have evaluated DL performance in AMD diagnosis. Leng et al [9] reported a pooled sensitivity of 94% and specificity of 97% for convolutional neural network–based algorithms, while Chen et al [10] highlighted the overall superiority of artificial intelligence (AI) over retinal specialists. However, these prior reviews have notable limitations: they did not stratify human-AI comparisons by clinician experience level, used conventional bias assessment tools rather than the recently developed Prediction model Risk Of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI) instrument [11], and did not separately evaluate the clinically critical task of differentiating wAMD from dAMD. Moreover, the rapid advancement of DL architectures, particularly vision transformers, necessitates an updated quantitative synthesis incorporating the latest evidence.

Importantly, our review was designed to address several evidence gaps that were not fully covered in previous meta-analyses. First, instead of evaluating DL algorithms in isolation, we directly compared DL performance with ophthalmologists and further stratified these comparisons by clinician experience level. This is clinically relevant because screening and referral decisions are often made by clinicians with different levels of expertise. Second, we separately evaluated the classification of wAMD versus dAMD, a task with immediate therapeutic implications because delayed recognition of wAMD may postpone anti–vascular endothelial growth factor treatment. Third, we incorporated PROBAST+AI, a recently developed tool tailored to prediction models using AI, thereby providing a more AI-specific assessment of bias than conventional quality appraisal tools [12]. Fourth, we examined prediction intervals (PIs), validation strategy, imaging modality, and other sources of heterogeneity to move beyond average pooled performance and assess the likely robustness of DL algorithms across clinical settings. These features make the current review not only an update of the evidence base, but also a more deployment-oriented synthesis of the clinical value and limitations of DL for AMD image classification.

Therefore, this systematic review and meta-analysis addressed these clinically relevant and deployment-oriented evidence gaps. Its objective was to evaluate the diagnostic performance of DL algorithms compared with ophthalmologists of varying experience levels for detecting AMD and differentiating its subtypes (wAMD vs dAMD), and to assess potential factors influencing DL diagnostic performance through subgroup analyses and meta-regressions.

In relation to our objective of comparing DL algorithms with ophthalmologists and identifying factors that influence diagnostic performance, the main finding is that DL models showed strong pooled performance for both AMD detection and wAMD versus dAMD classification, but the strength of this evidence differs across comparisons. The clinician comparisons suggest a possible role for DL as a consistent decision-support baseline, yet the sparse junior-ophthalmologist data and wide between-study variability mean that these findings should be interpreted as hypothesis-generating rather than definitive for deployment.

Our meta-analysis reveals a deployment-relevant pattern. Within the analyzed datasets, DL algorithms demonstrated significantly higher pooled sensitivity and accuracy compared to the available metrics for senior ophthalmologists in distinguishing AMD from normal controls; however, no significant differences were observed between DL and junior ophthalmologists across any diagnostic metrics. The higher pooled performance of DL suggests that these models possess an enhanced capability to detect subtle, pixel-level morphological changes and nonlinear feature interactions, such as early exudative signs, that may elude the visual inspection of even the most experienced clinicians [20,24,36]. Conversely, the lower sensitivity observed in senior ophthalmologists appears to be behaviorally driven by specific decision thresholds rather than skill deficiencies. As evidenced by the data from studies such as Matsuba et al [20], Bao et al [24], and Oliveira et al [36], senior experts exhibited a distinct preference for conservative diagnostic thresholds, achieving high specificity (summary 0.93) but notably reduced sensitivity (summary 0.75). This indicates a clinical preference for “rule-in” strategies to strictly avoid false positives, a constraint that DL algorithms do not possess. However, the unexpected parity between DL algorithms and junior practitioners should be interpreted with caution; this finding is attributable to data sparsity rather than clinical equivalence, as the analysis is restricted to only one study for AMD detection and two studies for wAMD classification, introducing a high risk of small-sample bias [20,24,36]. Similarly, the comparisons with senior ophthalmologists, while more robust, remain limited in number (three studies for AMD detection, two for wAMD classification), and the resulting estimates should be considered preliminary rather than definitive. This data-driven limitation, however, stands in contrast to the wAMD versus dAMD task, where sufficient head-to-head comparisons (two studies) enabled a stratified analysis, revealing distinct behavioral patterns across experience levels [20,24,36].

The very high AUC values observed in our analysis warrant careful interpretation and should not be equated with flawless real-world diagnostic performance. Restricted test distributions, curated image quality, internal validation, repeated use of public datasets, model selection based on the best-performing algorithm, and insufficiently documented patient-level splitting may inflate discrimination [25,38,43,46]. Potential data leakage cannot be excluded in studies that did not clearly separate patients, eyes, or images across training, validation, and test sets. Therefore, AUCs close to 1.0 indicate excellent discrimination within the analyzed datasets, not proof of flawless performance in prospective clinical workflows [47,48].

In the clinically critical task of differentiating wAMD from dAMD, automated DL systems demonstrated robustness across varying datasets. Our results indicated that the pooled metrics of DL algorithms not only showed higher specificity and accuracy compared to junior ophthalmologists but also indicated higher sensitivity and accuracy relative to senior ophthalmologists. Rather than framing this as DL mitigating an “experience gap” or correcting specific human errors, these findings suggest that DL algorithms offer a more consistent and objective diagnostic baseline that balances sensitivity and specificity. These findings advocate for a collaborative clinical paradigm: DL algorithms could serve as a triage filter to enhance specificity for primary care providers while functioning as a high-sensitivity “second reader” for specialists resolving equivocal wAMD cases [49].

Interestingly, our subgroup analysis highlights the relatively consistent performance of OCT-based models for automated AMD classification. OCT-based models significantly outperformed CFP-based approaches, driven by the capture of pathognomonic cross-sectional features—such as intraretinal fluid and pigment epithelial detachment—that are often obscured in 2D fundus photography [50]. Intriguingly, multimodal DL (OCT + CFP) did not significantly surpass standalone OCT models. This suggests a “saturation effect,” where the rich structural data of OCT capture the vast majority of diagnostic signals, rendering the incremental value of CFP marginal [35]. In practice, the underperformance of multimodal models relative to standalone OCT may also stem from feature redundancy and fusion noise [51]; when OCT and CFP capture substantially overlapping diagnostic information [43], their combination can paradoxically introduce variance through misaligned spatial features, registration errors, and conflicting feature representations, ultimately degrading rather than enhancing the decision boundary. From a translational perspective, this finding is pivotal; it implies that the computational cost and technical challenges of multimodal alignment (eg, fusion noise and registration errors) may currently outweigh the clinical benefits [51,52]. Therefore, unless fusion strategies are substantially optimized, OCT-based workflows currently appear to be a practical foundation for clinical deployment [35].

Building upon the baselines established by previous meta-analyses, this systematic review advances the understanding of DL in AMD diagnosis. In 2023, Leng et al [9] reported a pooled sensitivity of 94% and specificity of 97% for convolutional neural network algorithms. More recently, Chen et al [27] highlighted the superiority of AI over retinal specialists. Our analysis incorporates the latest studies using advanced architectures, such as Vision Transformers. This inclusion yields modestly higher pooled metrics, reflecting the field’s technological maturation [33]. Most significantly, this systematic review distinguishes itself through four methodological innovations that enhance clinical relevance: (1) a stratified comparison of AI versus ophthalmologists, explicitly differentiating by experience level; (2) the application of the PROBAST+AI tool for bias assessment, complemented by the GRADE framework; (3) a rigorous subgroup analysis by imaging modality (OCT, CFP, multimodal) to isolate technical performance drivers; and (4) a granular evaluation extending beyond binary detection to the specific classification of wAMD versus dAMD. Collectively, these advancements establish a more robust evidence base than prior reviews.

Heterogeneity is central to the interpretation of these findings [53]. Extreme between-study variability should not be treated only as a statistical descriptor; it indicates that pooled estimates may not transfer reliably to clinics with different devices, acquisition protocols, labeling rules, disease spectra, or patient populations [53-55]. By using a bivariate random-effects model and multivariable meta-regression, we identified that differences in validation strategies (internal vs external), database sources (open vs. private), and study settings (single-center vs. multicenter) significantly influence diagnostic performance. Studies relying solely on internal validation frequently reported inflated metrics, illustrating a generalization gap when models face domain shifts in image acquisition or demographics [50]. Similarly, single-center studies risk overfitting to specific center features arising from uniform protocols, whereas multicenter designs typically demonstrate greater robustness through exposure to diverse image qualities [50]. Consequently, our analysis suggests that database diversity and annotation quality are likely more critical determinants of generalizability than mere data accessibility. Furthermore, specific outliers in the bivariate box plot (Skevas et al [21], Celebi et al [26], Grassmann et al [31], El-Den et al [29], Le et al [33], Wang et al [42], and Yoo et al [43]) highlight how methodological divergences, such as algorithm architecture and data curation strategies, can materially affect performance. This indicates that future improvements in AI reliability will depend less on novel model architectures and more on the curation of diverse, multicenter external validation datasets. The substantial heterogeneity observed in this systematic review and meta-analysis warrants careful interpretation. Rather than reflecting routine statistical noise, this level of heterogeneity signals the pooling of fundamentally diverse data sources. Specifically, the included studies used different imaging hardware (eg, Heidelberg Spectralis, Topcon, Zeiss Cirrus OCT devices; various fundus camera systems), acquisition protocols (varying image resolutions, fields of view, and scan patterns), and ground-truth labeling methodologies (ranging from consensus grading by multiple retinal specialists to single-expert annotation or semi-automated classification systems) [32,35,45]. These technical and methodological differences fundamentally influence the feature space available to DL algorithms and likely account for much of the observed heterogeneity. The PIs (eg, sensitivity: 0.95‐0.99; specificity: 0.95‐0.99 for distinguishing AMD from normal retinas; sensitivity: 0.89‐0.97; specificity: 0.92‐0.97 for classifying wAMD vs dAMD) further underscore that the average pooled performance, while encouraging, may not be representative of performance in any individual deployment setting. This finding has important implications for clinical deployment: site-specific validation using local imaging equipment and patient populations remains essential before implementing any DL-based AMD screening system.

These heterogeneity findings provide concrete guidance for future study design [56]. Investigators should prioritize external validation on datasets from institutions and populations distinct from the training data [53], use multicenter designs incorporating diverse imaging devices and acquisition protocols, report patient-level separation between training and testing data, and stratify performance by imaging device, acquisition protocol, labeling method, and patient demographics [56]. Because PIs indicate that local performance may differ from pooled estimates, summary metrics alone should not be used as a deployment decision rule [54].

Translating these findings into practice, DL algorithms exhibit the potential to augment the diagnostic workflow rather than replace it. The superior accuracy of DL in classifying wAMD versus dAMD suggests potential use in resource-limited settings or tele-ophthalmology screening. However, considering current algorithms are primarily trained on isolated OCT or CFP images, they often lack integration with other imaging modalities or clinical parameters; future models should therefore evaluate multimodal imaging and patient clinical contexts to emulate comprehensive diagnoses [57]. Beyond these technical and clinical considerations, significant implementation barriers persist, including the scarcity of expert-annotated data, regulatory hurdles, and technical challenges regarding data availability, model interpretability, transparency, and generalization capability [36,52,58]. Advances in few-shot learning, self-supervised models, and centralized platforms may support a more integrated AI ecosystem, requiring sustained multidisciplinary efforts to optimize AI safety and support safe clinical practice [52,58].

Beyond diagnostic performance metrics, the successful clinical translation of DL algorithms requires addressing practical implementation challenges. These include seamless integration into existing electronic health record systems and ophthalmic imaging workflows, real-time processing capabilities compatible with clinical time constraints, and intuitive user interfaces that present AI-generated results in a manner that supports rather than disrupts clinical decision-making [46]. Furthermore, clinician trust and acceptance—shaped by model interpretability, transparency of AI reasoning, and consistent performance across diverse clinical scenarios—are prerequisites for successful adoption [23,24]. Future validation of DL tools must therefore extend beyond accuracy benchmarks to encompass usability studies, clinician acceptance evaluations, and workflow efficiency assessments in real-world clinical settings [41].

Our findings should be interpreted considering several limitations. First, the predominance of retrospective study designs (26 of 28 included studies) represents a fundamental limitation that must be carefully considered when interpreting the strong pooled performance metrics. Retrospective datasets are typically curated from clinical archives, which may systematically exclude poor-quality images, atypical presentations, and diagnostically challenging cases that are routinely encountered in prospective clinical workflows. This selection inherently inflates the apparent diagnostic performance and limits the generalizability of our findings to real-world screening and clinical deployment settings [48]. Second, to address potential patient overlap and maintain statistical independence, we extracted performance metrics exclusively from the primary AI algorithm within each study, omitting data from suboptimal models. While methodologically sound for meta-analysis, this approach inherently reflects a “best-case scenario” that likely inflates the pooled performance estimates compared to average algorithmic performance. This reporting bias is an inherent limitation of the current DL literature in ophthalmology and should be carefully considered by clinicians and policymakers when interpreting these results for clinical implementation decisions [12]. Future research should therefore granularly evaluate performance variances across different algorithmic architectures, including less optimal models, to ensure a more balanced and realistic assessment of the DL landscape. Third, direct head-to-head comparisons between DL algorithms and ophthalmologists were small, particularly for the AMD versus normal task where only one study provided junior ophthalmologist data, limiting the statistical power of these specific subgroups [50]. Future research must prioritize prospective, multicenter trials with prespecified human comparison arms to definitively validate these retrospective results [12].

In conclusion, this systematic review suggests that, compared with ophthalmologists, DL algorithms demonstrate superior and more balanced diagnostic performance for AMD image classification, providing a consistent decision-support baseline that mitigates the threshold-dependent trade-offs observed in human graders. However, these relative-performance findings remain preliminary because head-to-head evidence is sparse, especially for junior ophthalmologists, and because wide PIs, high heterogeneity, retrospective designs, and possible inflation from restricted datasets, internal validation, or leakage limit clinical transportability. DL systems should therefore be locally calibrated and prospectively validated as triage adjuncts rather than autonomous replacements. Before implementation, prospective multicenter studies should test representative patients, use strict patient-level external validation [56], include prespecified human comparison arms [59], and evaluate workflow integration, interpretability, and safety [60].