- Home
- About the Journal
- Peer Review
- Editorial Board
- For Authors
- Reviewer Recognition
- Archive
- Contact
- Impressum
- EWG e.V.
Cite as: Archiv EuroMedica. 2026. 16; 2. DOI 10.35630/2026/16/Iss.2.21
,
Zuzanna
Dynowska1 , ,
Zuzanna
Muszkiet3 , ,
Adam
Janota3 , ,
Michał
Bar2 , Uveal melanoma is the most common primary intraocular malignancy in adults and remains an important cause of metastatic mortality despite advances in ocular imaging and local therapy. Early detection is clinically important because it enables timely referral to ocular oncology services and supports eye preserving treatment strategies. However, small melanomas may resemble benign choroidal nevi, and their differentiation often requires interpretation of multimodal ophthalmic imaging. Artificial intelligence, particularly deep learning, has recently expanded in ophthalmic image analysis and offers potential tools for automated detection, classification, and risk assessment of melanocytic lesions.
To synthesize current scientific evidence on the use of artificial intelligence for analysis of multimodal ophthalmic imaging in the early detection of uveal melanoma and its differentiation from choroidal nevi, and to identify major methodological and clinical factors influencing translation into clinical practice.
This narrative review was conducted using a structured literature search of PubMed MEDLINE and cross referenced sources to identify relevant publications from 2017 to 2026. Search terms included uveal melanoma, choroidal nevus, artificial intelligence, deep learning, machine learning, and ophthalmic imaging modalities such as fundus photography, ultra widefield imaging, optical coherence tomography, OCT angiography, ultrasonography, and magnetic resonance imaging. The evidence was analyzed using thematic narrative synthesis, and reporting quality and potential bias were interpreted in the context of contemporary methodological guidance for artificial intelligence studies.
Available studies show that deep learning algorithms can differentiate choroidal nevi from small uveal melanomas using fundus photographs and that multimodal models integrating ultra widefield fundus imaging with ocular ultrasonography may improve classification performance compared with single modality approaches. Artificial intelligence models have also been applied to the detection of risk features of choroidal nevi and to radiomics analysis of magnetic resonance imaging for prognostic assessment. However, the current evidence base remains limited by relatively small datasets, heterogeneous diagnostic reference standards, class imbalance, and limited external validation.
Artificial intelligence applied to multimodal ophthalmic imaging demonstrates potential as a decision support and triage tool for earlier and more consistent detection of uveal melanoma and high risk melanocytic lesions. Further research should focus on multicenter datasets, standardized diagnostic labeling, rigorous external validation, and prospective clinical evaluation aligned with contemporary reporting and risk of bias frameworks.
Keywords: Uveal melanoma; choroidal nevus; artificial intelligence; deep learning; multimodal imaging; fundus photography; ultrasonography; optical coherence tomography; magnetic resonance imaging; radiomics.
Uveal melanoma (UM) is the most common primary intraocular malignancy in adults and arises from melanocytes of the uveal tract (choroid, ciliary body, iris) [1,2]. Although many patients can be treated successfully with eye-preserving local therapies, approximately half develop metastatic disease, most frequently to the liver, with historically poor survival once metastasis occurs [1,2]. The biology of UM differs substantially from cutaneous melanoma, including characteristic driver mutations (e.g., GNAQ/GNA11) and distinct molecular subtypes, which influence prognosis and metastatic risk [3–6]. Recent systemic therapy advances, including tebentafusp for HLA-A*02:01-positive metastatic UM, provide the first proven overall survival benefit in a phase 3 setting, but do not eliminate the high unmet need across the disease continuum [7–9].
From a clinical standpoint, “early detection” in UM has two pragmatic meanings. First, it encompasses earlier identification and referral of malignant lesions when they are small, potentially enabling globe-sparing treatment, reducing complications, and shortening time to definitive management. Second, it includes earlier discrimination of small melanomas from benign choroidal nevi, which are common and often incidentally detected. Differentiating a small melanoma from a nevus is a classic diagnostic challenge requiring synthesis across multimodal imaging (color fundus photography, ultra-widefield imaging, autofluorescence, OCT, OCT angiography, and ultrasonography), supplemented by clinical features and longitudinal change [10–15].
In real-world practice, the diagnostic pathway is heterogeneous. In many systems, initial detection occurs in general ophthalmology or optometry, where access to comprehensive ocular oncology expertise and multimodal imaging may be limited. Referral decisions may therefore depend on subjective assessment, variable imaging quality, and inconsistent follow-up intervals. Diagnostic delay is clinically meaningful: delayed ocular oncology evaluation can lead to treatment at larger tumor size, with potential impacts on vision outcomes and metastatic risk [12,13]. These realities motivate interest in scalable decision-support and triage tools.
Artificial intelligence (AI) has rapidly matured in ophthalmology, driven by large image datasets, standardized imaging modalities, and a track record of clinically validated deep learning systems for retinal disease screening and referral [16–21]. Parallel advances in medical imaging AI, explainability methods, and deployment considerations support the feasibility of AI-based tools for ocular oncology, including UM and its mimickers [22-28]. Importantly, the promise of AI in UM hinges on multimodal imaging: many clinically discriminative features (e.g., subretinal fluid, orange pigment surrogates, thickness/reflectivity on ultrasound) are not fully captured by a single modality, and reference standard diagnosis often relies on multimodal consensus by ocular oncologists [10–15]. Thus, multimodal AI is an intuitively appropriate strategy.
Uveal melanoma remains a clinically significant oncologic problem in ophthalmology because it combines relatively low incidence with a high risk of metastatic spread and disease related mortality [1,2]. The biological behavior of this tumor differs from that of cutaneous melanoma and is associated with characteristic molecular alterations, including mutations in GNAQ and GNA11 and distinct molecular subtypes that influence metastatic potential and clinical prognosis [3–6]. Although local treatments allow preservation of the eye in many patients, survival outcomes remain strongly dependent on the development of metastatic disease.
An important clinical challenge is the early identification of malignant lesions among the large number of benign choroidal nevi detected during routine ophthalmic examinations. Small melanomas may initially present with subtle clinical features and therefore require careful assessment using several complementary imaging techniques. Modern ophthalmic diagnostics relies on multimodal imaging, including color fundus photography, ultra widefield imaging, optical coherence tomography, OCT angiography, and ultrasonography, which together provide information about lesion morphology, retinal alterations, vascular changes, and tumor thickness [10–15].
In routine clinical practice, the first detection of suspicious choroidal lesions often occurs outside specialized ocular oncology centers, particularly in general ophthalmology or optometry settings. Limited access to multimodal imaging and subspecialty expertise may lead to variability in clinical interpretation and delayed referral for expert evaluation. In this context, the development of artificial intelligence methods for medical image analysis has attracted increasing attention. Recent advances in deep learning for ophthalmic imaging create opportunities for automated image interpretation systems that could assist clinicians in identifying suspicious lesions and supporting referral decisions in patients with suspected uveal melanoma [16–21].
The scientific novelty of this article lies in the structured synthesis of contemporary research on the application of artificial intelligence for the analysis of multimodal ophthalmic imaging in the early detection of uveal melanoma and its differentiation from choroidal nevi. The article integrates evidence across several imaging modalities, including color fundus photography, ultra widefield imaging, optical coherence tomography, OCT angiography, ultrasonography, and magnetic resonance imaging, which represent complementary sources of diagnostic information in clinical ophthalmology [10–15,39–41]. The review analyzes existing deep learning models for the classification of choroidal lesions, including algorithms based on fundus photographs, multimodal models combining ultra widefield imaging and ultrasound, and models designed to identify risk factors for malignant transformation of choroidal nevi [35–38]. The article also discusses emerging research directions, including the use of magnetic resonance imaging based radiomics for prognostic modeling in uveal melanoma [39]. In this way, the review provides an integrated overview of current research on artificial intelligence in ocular oncology and highlights important methodological limitations of existing studies.
The aim of this review article is to summarize and analyze current scientific evidence on the application of artificial intelligence methods for the analysis of multimodal ophthalmic imaging for the early detection of uveal melanoma and its differentiation from benign choroidal nevi [35–38].
Research objectives
This work is a narrative review with a structured literature search. The search and selection of studies followed established principles of transparent evidence identification and were informed by PRISMA 2020 recommendations where applicable [29]. Because the included studies show substantial heterogeneity in AI model design, datasets, imaging modalities, and reported outcomes, quantitative meta analysis was not appropriate. The evidence is therefore synthesized thematically with emphasis on clinical relevance and methodological interpretability.
Eligibility criteria
Inclusion criteria
We searched PubMed/MEDLINE for relevant literature. In addition, the reference lists of relevant articles and authoritative reviews were screened to identify additional eligible studies. Search terms combined descriptors related to uveal or choroidal melanoma with artificial intelligence terms and imaging modalities. Representative search strings included: (“uveal melanoma” OR “choroidal melanoma” OR “choroidal nevus”) AND (“artificial intelligence” OR “deep learning” OR “machine learning”) AND (fundus OR “ultra-widefield” OR OCT OR “optical coherence tomography” OR ultrasonography OR ultrasound OR MRI OR radiomics OR multimodal).
The search aimed to identify representative studies illustrating the main approaches to the application of artificial intelligence in ophthalmic imaging for uveal melanoma detection and risk assessment.
Titles and abstracts were screened for relevance. Full texts were then reviewed to confirm eligibility. Extracted items included study design, population size, imaging modalities, AI architecture (when reported), reference standard definition, validation strategy (internal vs external), performance metrics (AUC, sensitivity, specificity), and key limitations.
Given the nature of AI prediction models, we structured critical appraisal around contemporary reporting and bias guidance for AI and prediction modeling. We used TRIPOD+AI for reporting expectations in prediction model studies [31], and PROBAST+AI principles to frame risk-of-bias and applicability concerns [32]. For narrative synthesis quality, we considered SANRA domains as an organizing tool for completeness and transparency [30,33]. For clinical evaluation studies involving AI interventions, we referenced CONSORT-AI requirements conceptually [34]. These tools were used qualitatively to guide critique rather than as a formal scoring exercise, because many included studies were development-stage rather than prospective clinical trials.
The literature on AI for UM detection and classification has expanded notably in the past 2–3 years, shifting from proof-of-concept single-modality models to more clinically aligned multimodal approaches. The evidence can be grouped into five archetypes:
We additionally considered foundational UM biology and clinical risk factor literature to contextualize AI targets and the clinical plausibility of imaging biomarkers [1–6,10–15].
UM is often detected as an intraocular mass on clinical examination or incidental imaging. In practice, early lesions may be small, minimally elevated, or partially obscured by media opacities, and the diagnostic question frequently becomes whether a lesion is a benign nevus warranting observation or a melanoma requiring urgent referral and treatment.
Multimodal imaging is essential because different modalities capture distinct components of the lesion and its microenvironment:
Clinically, the diagnostic “reference standard” for many early lesions is not histopathology (which is rarely obtained in small tumors), but expert consensus supported by multimodal imaging and longitudinal behavior (growth, development of risk features) [10–15]. This has immediate implications for AI: models must be trained with reference standards that reflect real-world diagnostic ground truth, and validation must anticipate that “truth” is often a composite construct rather than a single test.
The key ophthalmic imaging modalities relevant to early uveal melanoma detection and the corresponding AI-relevant biomarkers and analytical tasks are summarized in Table 1.
Table 1. Multimodal imaging modalities and AI-relevant biomarkers for early uveal melanoma detection
| Imaging modality | Key lesion features detectable | AI-relevant biomarkers | Potential AI tasks | Clinical relevance | References |
| Color fundus photography | Lesion pigmentation, borders, drusen, RPE alterations | Shape, color heterogeneity, border irregularity | Image classification (nevus vs melanoma), lesion detection | First-line imaging and screening | [10–13,35,36] |
| Ultra-widefield (UWF) imaging | Peripheral lesion visualization, lesion extent | Peripheral morphology, size estimation | Classification, lesion localization | Improved visualization of peripheral tumors | [37] |
| Fundus autofluorescence (FAF) | Lipofuscin accumulation (“orange pigment”) | Autofluorescence intensity patterns | Risk-feature detection | Indicator of malignant potential | [10–15] |
| Optical coherence tomography (OCT) | Subretinal fluid, retinal layer disruption | Fluid segmentation, photoreceptor alterations | Segmentation, biomarker detection | Detection of secondary retinal changes | [14,40] |
| Optical coherence tomography angiography (OCTA) | Microvascular changes, vascular density alterations | Flow signal patterns, vascular networks | Vascular feature analysis, longitudinal monitoring | Monitoring tumor-related vascular changes | [40,41] |
| Ocular ultrasonography | Tumor thickness, internal reflectivity | Thickness measurements, acoustic reflectivity | Classification, segmentation, growth monitoring | Core modality for melanoma confirmation | [10–13,37] |
| Magnetic resonance imaging (MRI) | Tumor extent, extraocular extension | Radiomic texture features | Prognostic modeling | Systemic staging and prognostic assessment | [15,39] |
Several recent studies demonstrate that deep learning can meaningfully differentiate choroidal nevi from small melanomas using fundus photographs alone.
A multicenter Ophthalmology Science study developed and validated a deep learning algorithm to differentiate choroidal nevi from small melanoma using wide field and standard field fundus photographs, with external clinic testing and comparison against ophthalmologist performance [38]. Key strengths include multicenter image sources and external testing, as well as clinically meaningful endpoints aligned with referral decisions. The study underscores that AI can achieve high sensitivity at moderate specificity, suggesting a potential role as a triage tool where false negatives must be minimized.
A Journal of Clinical Medicine study evaluated deep learning to distinguish highly malignant UM from benign choroidal nevi using fundus photographs, emphasizing human–machine interaction (HMI) and exploring a workflow where the model supports clinician decision-making rather than replacing it [35]. This is clinically relevant because early deployment in ocular oncology is likely to occur as decision-support, where clinicians retain responsibility and the system is designed for interpretability and error checking.
These studies collectively suggest that fundus-based AI can:
A major advance in clinical realism is the emergence of multimodal AI models that integrate fundus photography with ultrasonography, reflecting how ocular oncologists evaluate lesions. An attention-based multimodal deep learning study in Ophthalmology Science integrated ultra-widefield fundus images and B-scan ultrasound to classify UM vs choroidal nevi [37]. The design aligns strongly with real-world diagnostics: ultrasound contributes thickness and reflectivity cues, while fundus imaging provides surface-level morphological context and lesion extent.
Key implications:
However, the clinical translation of multimodal AI raises additional implementation challenges:
Thus, multimodal AI may be best positioned initially at the level of ophthalmology clinics and ocular oncology referral networks, where both imaging streams exist and quality control can be enforced.
A complementary approach is to predict clinically meaningful risk factors using limited input modalities, reducing dependence on multimodal equipment at early stages. An AJO International study developed a deep learning model to classify choroidal melanoma risk factors based on color fundus photographs, explicitly motivated by the limited availability of multimodal imaging in resource-limited settings [38]. This approach is clinically attractive because it targets actionable intermediate phenotypes (e.g., risk features associated with malignant transformation), rather than forcing a binary “nevus vs melanoma” decision on a single modality.
Conceptually, this aligns with how ocular oncologists reason:
This approach also supports modular multimodal extension: a system could output fundus-derived risk features and later incorporate ultrasound or OCT-derived risk features when available, gradually improving certainty.
OCT-based characterization of small choroidal melanocytic lesions has been explored for years, including enhanced depth imaging OCT to visualize small melanomas and related choroidal structure [14]. More recently, narrative syntheses of OCT and OCTA applications in UM highlight potential diagnostic and monitoring roles, including detection of subretinal fluid, photoreceptor disruption, and treatment-related vascular changes [40]. A case-based OCTA report in plaque brachytherapy-related radiation retinopathy illustrates how OCTA captures longitudinal microvascular compromise after UM treatment, which may be relevant for AI monitoring and complication prediction rather than primary detection [41].
For AI, OCT/OCTA offers:
Yet the evidence for AI specifically targeting OCT/OCTA in UM remains less mature than fundus and ultrasound models. Key barriers include smaller labeled datasets, variability in scan protocols, and the need for robust annotation (fluid, RPE changes, choroidal boundaries). The near-term opportunity is likely in adjunctive tasks: fluid detection, tumor-related retinal change quantification, and prediction of radiation retinopathy severity rather than primary lesion classification.
MRI-based radiomics extends AI beyond detection to prognostic modeling. A Journal of Computer Assisted Tomography study developed an MRI-based radiomics model to predict disease-free survival in pretreatment UM, reporting an initial validation approach [39]. This is relevant because systemic risk stratification is increasingly important in UM care, especially as systemic treatments and surveillance strategies evolve.
Radiomics models face distinct challenges:
Nevertheless, the direction is consistent with broader oncology trends: multimodal AI may ultimately integrate ocular imaging, systemic imaging, and molecular predictors into unified risk models.
Across the most clinically proximal evidence, there is a consistent “performance signal” suggesting that deep learning can:
However, across studies, common constraints limit certainty about generalizability:
Table 2. Summary of key AI studies for uveal melanoma and choroidal nevus differentiation (2019–2026)
| Study | Imaging modality | Dataset / study population | AI approach | Reference standard | Validation | Key findings | References |
| Sabazade et al., 2024 | Fundus photography | Multicenter dataset of choroidal nevi and small melanoma | Deep learning CNN | Expert diagnosis with multimodal imaging and follow-up | External validation | High discrimination between nevi and melanoma | [33] |
| Hoffmann et al., 2024 | Fundus photography | Fundus images of benign and malignant lesions | Deep learning classifier with human–machine interaction | Clinical diagnosis by ocular oncology specialists | Internal validation | Accurate differentiation with clinically relevant sensitivity | [34] |
| Dadzie et al., 2025 | Ultra-widefield imaging + ultrasound | Multimodal imaging dataset of choroidal lesions | Attention-based multimodal deep learning | Multimodal expert diagnosis | Independent test set | Multimodal models outperform single-modality models | [35] |
| Dalvin et al., 2025 | Color fundus photography | Dataset of melanocytic lesions with known risk features | Deep learning risk-factor detection model | Clinical risk criteria for malignant transformation | Internal validation | Prediction of melanoma risk features | [36] |
| Su et al., 2023 | MRI radiomics | Pretreatment MRI of UM patients | Radiomics-based machine learning model | Clinical outcomes and follow-up | Internal validation | Prediction of disease-free survival | [38] |
| Shields et al., 2019 | Fundus + clinical imaging | Large cohort of choroidal nevi with longitudinal follow-up | Quantitative imaging analysis | Longitudinal malignant transformation | Observational dataset | Identification of imaging risk features | [37] |
The most plausible near-term clinical role for AI in early UM detection is triage and decision support. This is because:
Fundus-only models can function as an “early warning system,” flagging lesions that merit prompt multimodal imaging or specialist referral [35,36]. Multimodal models integrating ultra-widefield fundus and ultrasound can support more specialized clinics where both modalities are available, potentially improving accuracy and consistency in borderline cases [39]. Risk-factor detection models can provide interpretable outputs aligned with clinician reasoning and could be used to structure referrals and follow-up schedules [38].
A central challenge in ocular oncology AI is defining the ground truth. For many small lesions, histopathology is unavailable; instead, “truth” is inferred from:
This creates label noise and temporal leakage risks. For example, if a lesion is labeled as “nevus” because it did not progress over 5 years, the model may inadvertently learn site-specific follow-up patterns or imaging artifacts associated with surveillance rather than biology. Conversely, lesions treated early may not have prolonged follow-up, and their labels may rely on pre-treatment expert judgment.
Practical mitigation strategies include:
UM is relatively rare compared with common retinal diseases. This scarcity drives several issues:
Multimodal datasets are even harder to assemble, as they require paired imaging across modalities and consistent clinical labeling [37]. This highlights the importance of multicenter consortia, standardized protocols, and prospective dataset building in ocular oncology.
Generalizability is the major barrier between promising results and clinical deployment. Domain shift can be severe in ophthalmic imaging:
Medical imaging AI literature emphasizes that models trained on one domain may degrade on unseen domains unless robust validation and domain generalization strategies are used [28]. In ocular oncology, this argues for:
For high-stakes oncology decisions, clinicians require interpretability. While saliency methods such as Grad-CAM provide visual heatmaps that can be used to assess whether the model attends to plausible lesion regions, these methods can be misleading if not carefully validated [25]. Complementary explainability frameworks (e.g., SHAP for structured models) can support interpretation when the model includes engineered features or multimodal fusion components [26].
In practice, explainability should be used as:
A key future direction is “clinically interpretable outputs” rather than purely post-hoc explanations. Risk-factor detection models exemplify this approach by outputting clinically recognized features instead of opaque scores [38].
Many AI studies historically suffer from incomplete reporting, unclear validation, and optimistic performance estimates. Modern reporting guidance directly addresses these gaps. TRIPOD+AI provides updated guidance for reporting prediction models using regression or machine learning, emphasizing transparency in data handling, model development, validation, and intended use [31]. PROBAST+AI provides a structured approach to assess risk of bias and applicability, including domains related to participants, predictors, outcomes, and analysis [32]. CONSORT-AI highlights what is required when AI interventions are evaluated in trials, including the need to describe human–AI interaction and handling of inputs/outputs [34]. These frameworks are increasingly essential if ocular oncology AI is to transition from retrospective development to prospective clinical validation.
Specific recurrent risk-of-bias patterns in UM imaging AI include:
The field should move toward pre-registered analysis plans, standardized outcome definitions, and prospective evaluation, even if initial deployment is limited to triage settings.
High AUC does not guarantee clinical benefit. For triage, the most relevant metrics are:
Given the potential harms of delayed melanoma diagnosis, AI tools should likely prioritize sensitivity. However, high false positives may overwhelm specialist services and erode trust. Therefore, workflow design matters: AI may be most useful when it triggers a structured next step (repeat imaging, ultrasound acquisition, expedited referral) rather than a single binary “malignant” label.
As systemic therapies improve (e.g., tebentafusp demonstrating survival benefit in metastatic UM) [9], the value of early detection and accurate risk stratification may increase: earlier diagnosis enables timely local treatment and may support earlier systemic surveillance and trial enrollment for high-risk molecular or imaging phenotypes. Multimodal AI could eventually integrate ocular imaging (lesion phenotype), systemic imaging (e.g., liver surveillance), and molecular prognostic markers into a unified risk framework, consistent with broader oncology directions.
A feasible translation roadmap includes:
Phase 1: Retrospective multicenter validation
Phase 2: Prospective silent-mode evaluation
Phase 3: Clinical impact evaluation
Phase 4: Post-deployment monitoring
AI for early uveal melanoma detection is evolving toward clinically relevant multimodal approaches. Based on the analyzed literature, the following conclusions can be drawn.
Under these conditions, artificial intelligence supported multimodal imaging may improve referral accuracy, reduce diagnostic delay, and support early detection and risk stratification in uveal melanoma.
The authors acknowledge the use of ChatGPT (OpenAI) for assistance in drafting and language editing of early manuscript versions. All sections were subsequently reviewed, revised and approved by the authors, who take full responsibility for the scientific content and conclusions.
The authors declare no conflicts of interest.
|
||
download article (pdf)