Current Topics on the Integration of Artificial Intelligence in the Histopathological and Molecular Diagnosis of Uveal Melanoma

Abstract

Background: This review examines the expanding influence of artificial intelligence (AI) in the detection and management of uveal melanoma (UM). Methods: This work delves into the application of AI technologies such as machine learning (ML), deep learning (DL), and convolutional neural networks (CNNs) in various diagnostic procedures, molecular profiling, and predictive analysis. Results: The discussion underscores AI’s potential to enhance diagnostic precision and efficiency. Particular focus is placed on its role in histopathological assessments of UM, where algorithms facilitate the analysis of whole-slide images (WSIs). AI contributes to more accurate tumor classification, assists in planning treatments, and improves the prediction of the prognostic indicators and molecular characteristics of the tumor. Conclusions: Despite these promising developments, this review acknowledges existing hurdles to AI implementation, including issues with data standardization and the interpretability of AI models. It emphasizes the need for further research to fully integrate AI into clinical workflows, ultimately aiming to improve patient care and outcomes.

1. Introduction

Uveal melanoma (UM) is the most frequent primary intraocular malignancy in adults. The most common site is the choroid (85–90%), whereas the iris and ciliary body represent rare locations (9–15%). Ciliary body melanomas are those with the poorest prognosis, whereas UMs arising in the iris generally have a better prognosis because they are discovered at an early stage. Around half of patients with UM develop metastasis, with the liver representing the most common metastatic site, resulting in death within a year. Patients with metastatic disease have a median survival of 4–15 months [1].

Ocular management for UM aims to retain the eye-ball and visual function. Enucleation used to be the preferred treatment but has now been replaced with a combination of phototherapy, radiotherapy, and local tumor resection. Different features, such as tumor diameter and location, the presence of vitreous hemorrhage, retinal detachment, or retinal involvement, have an impact on treatment decisions. Unfortunately, there is no definitive therapeutic strategy for metastatic UM [1].

In recent years, artificial intelligence (AI) has proven to be a valuable asset in oncology, including in its application in uveal melanoma (UM) [2]. This review aims to provide an overview of the most recent advancements in AI utilization for UM, focusing on studies published up to 2024. AI technologies such as machine learning (ML), deep learning (DL), and convolutional neural networks (CNNs) are pivotal in medical imaging and diagnostics. These tools are engineered to handle large datasets, recognize complex patterns, and generate insights that support clinical decision-making. In the context of UM, AI has been employed across various diagnostic domains, including image interpretation, molecular analysis, and predictive modeling.

2. Materials and Methods

2.1. Search Strategy

The authors conducted a comprehensive electronic search of the literature up to November 2024, utilizing the following databases to identify pertinent studies: PubMed/MEDLINE, Cochrane Library (including the Cochrane Database of Systematic Reviews and Cochrane Central Register of Controlled Trials—CENTRAL), Web of Science (Science and Social Science Citation Index), and Scopus. The search strategy combined both MeSH and non-MeSH terms, as detailed below.

(i) PubMed/MEDLINE: ((artificial intelligence [Title/Abstract]) OR (machine learning [Title/Abstract]) OR (Deep Learning [Title/Abstract]) OR (Convolutional Neural Networks [Title/Abstract]) OR (Histopathology [Title/Abstract]) OR (Molecular Diagnosis [Title/Abstract]) OR (Genomics [Title/Abstract])) AND (Uveal Melanoma [Title/Abstract]). Filters applied: English language. (ii) Cochrane Library: “artificial intelligence” in Title Abstract Keyword AND “uveal melanoma” in Title Abstract Keyword (including word variations). Language: English. (iii) Web of Science: “artificial intelligence” (Topic) AND “uveal melanoma” (Topic) OR “machine learning” (Topic) AND “uveal melanoma” (Topic), restricted to English language. (iv) Scopus: (TITLE-ABS-KEY (artificial intelligence) AND TITLE-ABS-KEY (uveal melanoma)) OR (TITLE-ABS-KEY (deep learning) AND TITLE-ABS-KEY (uveal melanoma)) AND (LIMIT-TO (LANGUAGE, “English”)).

Additionally, the reference lists of relevant publications were manually screened to ensure the inclusion of any studies potentially missed during the electronic search.

2.2. Inclusion and Exclusion Criteria

The studies included in the review were published in the period from September 2008 to November 2024. Articles that did not focus on uveal melanoma or approached it without emphasizing the use of AI in histopathological or molecular diagnosis, or in patient outcomes, were excluded. The review also omitted comments, opinions, perspectives, guidelines, editorials, case reports, previously published systematic reviews or meta-analyses, and papers in languages other than English. Publications limited to abstracts or those providing minimal or non-informative content were also excluded.

The inclusion criteria prioritized articles written in English with full-text availability, particularly those offering comprehensive analyses or detailed discussions of AI applications in UM. These applications included diagnosis, prognostic prediction, clinical management, and insights into molecular characteristics. Specifically, the inclusion criteria were as follows:

(i)

Involve patients diagnosed with UM;

(ii)

Use AI for histopathological or molecular diagnosis of UM;

(iii)

Use AI in predicting patient outcomes associated with UM.

3. Results

Our search of the literature identified 24 significant studies focusing on the application of AI in diagnosis, molecular analysis, and outcome prediction for patients with uveal melanoma (UM) [2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27]. In the realm of histopathology, AI algorithms, particularly deep learning (DL), have been utilized to identify UM and differentiate it from benign tumors such as choroidal nevi. Despite these advancements, challenges such as limited dataset availability and uncertain clinical applicability persist [28,29]. From a molecular perspective, AI has been employed to examine gene expression profiles (GEPs) and methylation patterns in UM. However, the field faces obstacles, including the need for larger, more diverse datasets and additional validation of the findings. AI models have demonstrated potential in improving treatment outcome predictions and stratifying prognostic risks, particularly concerning metastatic potential. Nevertheless, issues such as data biases, insufficient sample sizes, and the necessity for more extensive studies continue to hinder broader application. Relevant information was systematically extracted and is presented narratively. Figure 1 summarizes the main applications of AI in UM diagnosis and prognosis.

Figure 1. The main applications of AI in UM are summarized.

4. Discussion

4.1. AI in Histopathological Diagnosis

The methods commonly used to diagnose UM assess the tumor’s diameter and thickness to classify it into four T categories (from T1 to T4). Further staging considers ciliary body involvement and extraocular extension, which help define progressively higher-risk categories for metastatic death (ranging from T1a to T4e) [30].

Following this, lymph node involvement and distant metastases are evaluated, allowing the T stages to be grouped into AJCC stages, with progressively worse prognoses (ranging from stage I to IV). Another tumor-related feature associated with high mortality is the presence of epithelioid cells compared to spindle cells [30]. Additionally, a high mitotic figure count—indicating the increased proliferative activity of tumor cells—has also been linked to a poorer prognosis [31]. Histological analysis can also be used to examine blood vessels within the tumor. Microvascular density, assessed by staining for vascular markers such as CD31, CD34, and von Willebrand factor, is predictive of a worse prognosis [32]. Moreover, UM may exhibit extravascular loops and networks, best visualized with periodic acid–Schiff staining and a dark green filter. These structures are associated with poor prognosis [33].

In routine diagnostic pathology, the immunohistochemical expression of S100 and SOX10 has traditionally been considered the most sensitive marker of melanocytic lineage/differentiation when dealing with metastases from an unknown primary tumor. However, most studies on S100 immunoreactivity in melanocytic neoplasms have primarily focused on CM. Some studies, however, suggest a significant lack of S100 immunoreactivity in metastatic UM, in stark contrast to the widespread positivity seen in both primary UM and metastatic CM [34].

One of the primary applications of ML in this field is prognostication in uveal melanoma (UM). For example, DL-based analysis of digital pathology images shows great potential, particularly in predicting the gene expression profile (GEP) of UM tumors. Looking ahead, we anticipate that more advanced DL techniques—such as generative adversarial networks (GANs), few-shot learning, and zero-shot learning—will be increasingly used to address the challenges posed by the rarity of ocular oncology cases [35].

A study by Tahmasebzadeh et al. concluded that artificial intelligence is an emerging tool in image analysis, enhancing diagnostic workflows from data and image processing to clinical decision-making. It contributes to more personalized treatment approaches, improves response prediction, and refines prognosis assessment [36].

AI application in the histopathological analysis of UM represents a promising frontier. UM is a rare yet aggressive cancer of the eye, and its accurate diagnosis and prognosis assessment are crucial for determining appropriate patient treatment and management [37]. The histopathological examination of tissue samples remains the gold standard for diagnosing and classifying UM. However, the complexity of histological data and the subjective nature of human interpretation pose significant challenges in ensuring accurate diagnoses. AI, in particular DL algorithms such as CNNs, has shown potential in overcoming these limitations by providing a more reliable, reproducible, and efficient method of analyzing histopathological images.

One of the key advantages of AI in this domain is its ability to achieve high diagnostic accuracy in distinguishing between benign choroidal nevi and malignant UMs. Studies have demonstrated that AI algorithms outperform traditional diagnostic methods in certain contexts. For instance, a study by Dadzie et al. [2] highlighted that AI models could identify small melanomas with superior sensitivity compared to both resident and consultant ophthalmologists. The algorithm achieved comparable sensitivity and specificity to ocular oncologists, proving that AI can be an effective tool in complex diagnostic scenarios, especially when differential diagnosis is challenging. Moreover, AI shows consistency in assessing the same image every time it is analyzed, without human interobserver variability, and this may represent a crucial advantage in ensuring reliable results across different healthcare settings.

In addition to diagnostic accuracy, AI’s ability to integrate multiple data sources, such as histopathological images, clinical features, and genetic information, has further enhanced its value.

Another notable advantage of AI in histopathology is its ability to reduce human error and subjectivity in interpretation. Histopathological evaluation is often prone to interobserver and intraobserver variability, especially when assessing small or ambiguous lesions. AI algorithms trained on large datasets can provide consistent results by focusing on objective image features, significantly decreasing the potential for misclassification or oversight. As Chandrabhatla et al. pointed out, the use of AI models can be particularly beneficial in routine screenings and early detection, as the system can analyze large volumes of data efficiently, helping clinicians save time and focus on more complex cases [3].

Furthermore, the ability of AI to automatically segment and highlight regions of interest within histopathological images is another strength. Sabazade et al. noted that AI’s segmentation capabilities help isolate key features such as tumor boundaries, areas of infiltration, and abnormal tissue structures, thereby improving diagnostic clarity [4]. This segmentation not only increases classification accuracy but also makes the model’s decision-making process more interpretable, providing clinicians with insights into the specific features that contributed to the diagnosis. This transparency is critical for fostering trust between healthcare providers and the AI system, as it enables clinicians to better understand and validate AI-driven predictions.

However, despite these significant advantages, there are notable limitations that must be addressed for AI to reach its full potential in the histopathological analysis of UM. One of the most pressing challenges is the quality and diversity of the datasets used to train AI models. As highlighted by Wan et al., the accuracy of AI models is heavily dependent on the quality of the data they are trained on [5]. Small or non-representative datasets may lead to overfitting, where the model becomes highly specialized to the training data but fails to generalize well to new or unseen cases [5]. The rarity of UM further exacerbates this issue, as limited samples from diverse populations may hinder the model’s ability to detect atypical lesions or those with unusual presentations.

Moreover, while CNNs have demonstrated excellent performance in analyzing histopathological images, they are often considered “black-box” models. This means that the decision-making process of the AI system is not always transparent, which can be a barrier to clinical adoption. The lack of interpretability can make it difficult for clinicians to understand why the AI made a particular prediction or diagnosis, limiting their trust in the system. Techniques like saliency maps, as proposed by Zhang et al., could help address this issue by highlighting the key image features that influenced the model’s decision, thereby improving interpretability and fostering clinician confidence [6].

Additionally, AI’s ability to process only the data it has been trained on raises concerns regarding the generalizability of these models across different clinical settings. As noted by Sabazade et al., the algorithm they studied was trained on a limited set of images, which could potentially reduce its effectiveness when applied to different imaging systems or institutions with varying practices and equipment [4]. For AI to be successfully integrated into clinical practice, it must be trained on larger, more diverse datasets that include a wide range of image qualities and lesion types. This would ensure that the models are robust enough to perform well across different real-world scenarios.

Lastly, the implementation of AI in histopathology also requires validation through large-scale, prospective clinical studies. While retrospective studies have shown promising results, the true clinical value of AI can only be fully realized when it is tested in prospective trials with broader patient populations. Iddir et al. highlighted the importance of continued validation in diverse clinical settings to confirm the efficacy and reliability of AI algorithms in real-world applications [7].

In conclusion, while AI holds significant promise for improving the diagnosis, prognosis, and treatment of UM through histopathological analysis, challenges related to data quality, interpretability, and generalizability remain. Overcoming these limitations will require further research, improved dataset diversity, and the development of more transparent AI models to ensure their successful integration into routine clinical practice. The main advantages and limitations of AI applications in the histopathological diagnosis of UM are summarized in Table 1. AI, particularly DL and CNNs, enhances diagnostic accuracy and consistency in identifying UM, especially in distinguishing malignant tumors from benign nevi. Strengths include a reduction in human error, improved reproducibility, and the effective integration of multi-source data. However, limitations persist due to small, non-diverse datasets, interpretability challenges, and the lack of large-scale clinical validation.

Table 1. AI applications in UM histopathological diagnosis.

Study	Advantages	Limitations
Dadzie AK et al. (2024) [2]	AI provides consistent results, eliminating intraobserver and interobserver variability commonly seen among human specialists; DL, especially CNNs, has demonstrated significant capability in distinguishing UM from choroidal nevi, achieving high F1-scores, accuracy, and area under the curve (AUC) metrics in classification tasks	The performance of AI models heavily relies on the quality and diversity of training datasets. Limited or non-representative datasets can negatively affect results; lesions with unusual features or low-quality images (e.g., those affected by hemorrhage or retinal detachment) may not be accurately classified, limiting AI’s applicability
Chandrabhatla AS et al. (2023) [3]	The majority of ML research has demonstrated strong performance despite challenges like small datasets; tools developed with ML may assist in predicting metastatic spread for UM, which is a key area for improving patient outcomes	Limited sample sizes can lead to variability in algorithm performance and restrict generalizability
Sabazade S et al. (2024) [4]	The AI algorithm demonstrated superior sensitivity compared to resident and consultant ophthalmologists and matched the performance of ocular oncologists in differentiating small choroidal melanomas from nevi; the algorithm provides identical assessments for the same images, eliminating intraobserver and interobserver variability that can occur with human specialists	The algorithm’s accuracy depends heavily on the quality and precision of the ocular oncologists’ initial diagnoses and the representativeness of the dataset; a relatively small sample size was used for training and validation, which could limit the algorithm’s performance, particularly with atypical lesions not represented in the dataset
Wan Q et al. (2023) [5]	The algorithm achieved high accuracy (90%) in predicting vital status from histopathological images; AI identifies subtypes (Cluster1 and Cluster2) with distinct molecular profiles, mutations, and immune microenvironments, aiding in personalized therapeutic approaches	AI identified mutations and proposed biomarkers; although these results are promising, further validation of AI’s predictive value in independent clinical studies is essential for generalizing conclusions
Zhang H et al. (2021) [6]	AI was used in analyzing iris color for the potential diagnosis of uveal melanoma	There is a need for more comprehensive approaches, incorporating additional clinical and molecular data to improve diagnostic accuracy
Iddir SP et al. (2023) [7]	AI models demonstrated high accuracy in detecting risk factors for the malignant transformation of choroidal lesions using single ultra-widefield (UWF) or B-scan ultrasound (US) images; ML can streamline diagnosis by accurately analyzing risk factors such as subretinal fluid, orange pigment, and tumor thickness, which traditionally rely on subjective evaluation	The study involved a limited number of patients from a single institution, reducing the generalizability of the findings; small features like orange pigment and drusen are difficult to identify accurately, suggesting the need for alternative techniques such as image cropping
Kulbay M et al. (2024) [8]	AI models have shown high accuracy in distinguishing between benign choroidal nevi and malignant melanoma; combining histopathological data with clinical features enhances prediction accuracy	There is a need for larger datasets, better integration of imaging data, and the exploration of deep learning techniques
Hoffmann L et al. (2024) [9]	AI reduces dependency on expensive multimodal imaging, making it suitable for resource-limited settings; advanced techniques like attention mechanisms and larger datasets can enhance performance	Models were trained on small datasets with no external validation, affecting generalizability; misclassifications often occur in lesions with moderate or borderline features; real-world efficacy has not been confirmed through prospective clinical studies

4.2. AI in Molecular and Genomic Analysis

The BAP1 gene, situated on chromosome 3p21.1, encodes a nuclear ubiquitinase that plays a crucial role in the epigenetic regulation of chromatin. It is a key tumor suppressor, and its mutational inactivation is a pivotal step in the development of metastatic potential in uveal melanoma (UM) [10].

In a 2019 study, Sun et al. utilized a densely connected neural network to predict BAP1 expression from ophthalmic histopathology images, achieving a remarkable accuracy of 97% [11]. Similarly, Zhang et al. developed deep learning models to estimate nuclear BAP1 expression using hematoxylin and eosin (H&E)-stained histopathology slides, achieving an accuracy of 93% [12]. Gene expression profiling (GEP) is considered the best indicator to predict the survival of patients with UM, with two identified classes showing a significant difference in long-term survival (92-month reported survival is 95% and 31% for Class 1 and Class 2, respectively). Therefore, considering the importance of correctly assessing the GEP of UM, Liu et al. developed a DL model with a dual-attention feature extraction mechanism to directly predict the GEP from digital H&E-stained cytopathology images, achieving 91.7% sensitivity, specificity, and accuracy [13].

As the GEP represents a pivotal prognostic factor in UM, Akram et al. tried to predict molecular subclasses of UM based on H&E-stained sections through a CNN [14]. Their models correctly identified 14/20 BAP1^mut, 8/10 EIF1AX^mut, and 11/17 SF3B1^mut UM, suggesting that these models were able to predict the molecular signature of the tumor based on some morphological features.

Liu et al. conducted a study to develop a DL system that can determine patient survival using smeared cytologic aspirates from fine-needle aspiration biopsy (FNAB) samples, with the GEP as the reference standard [15]. This DL system predicted the GEP of the tumor in 15/20 cases (75%).

Genomic studies have revealed that uveal melanoma (UM) is driven by a small number of key driver events, resulting in four distinct molecular subtypes associated with varying metastatic risks. These subtypes are as follows: Class A (Disomy of chromosome 3 without hotspot mutations in SF3B1); Class B (Disomy of chromosome 3 with hotspot SF3B1 mutations); Class C (Monosomy of chromosome 3 with BAP1 mutations but no amplification of chromosome 8q); and Class D (Monosomy of chromosome 3 with BAP1 mutations, amplification of chromosome 8q, and an inflammatory infiltrate).

Metastatic risk is stratified by subtype, being low in Class A, intermediate in Class B, and high in Classes C and D [16].

In a 2024 study, Reggiani et al. utilized machine learning (ML) approaches to identify genes linked to copy number alterations (CNAs) and developed a model capable of distinguishing monosomic from disomic chromosome 3 samples, leveraging data from The Cancer Genome Atlas (TCGA) UM cohort [17]. Similarly, Xie et al. [18] conducted bioinformatics analyses of gene expression profiles in UM, including cases with metastasis, identifying several hub genes (e.g., FN1, APOB, F2, SERPINC1, SERPINA1, APOA1, FGG, PROC, ITIH2, VCAN, TFPI, CXCL8, CDH2, and HP) that may play crucial roles in the metastatic progression of UM.

Another interesting aspect to consider for predicting the clinical behavior of UM is the methylation profile of the tumor. In this regard, Hou et al. used ML models to detect DNA methylation signatures in UM, which may be of prognostic relevance [19]. They identified two risk groups, with significantly different prognosis in terms of survival.

The main advantages and limitations of AI applications to assess the GEP and molecular features of UM are summarized in Table 2. AI applications in this field have shown strong performance using routine histological images. These tools offer non-invasive, rapid molecular profiling potential. Limitations include reliance on small datasets, a lack of population diversity, and the need for broader validation across centers and staining techniques.

Table 2. AI applications to assess UM molecular features.

Study	Advantages	Limitations
Sun et al. (2019) [11]	Good accuracy in predicting BAP1 expression based on histopathology images through DL models	Histopathology slides from two centers were used, which may not adequately capture the full spectrum of nuclear BAP1 expression in UM or account for interlaboratory variability. A significant limitation for model training was the small sample size and the uneven distribution of the four categories (positive, negative, excluded, and blurred). Additionally, a loss of BAP1 expression can occur in a notable subset of cases without a corresponding BAP1 mutation, and the reverse is also possible, potentially creating a misleading impression of normal protein levels
Zhang et al. (2020) [12]	Good accuracy in predicting nuclear BAP1 expression through DL models on H&E-stained histopathology slides	Patches from the same slide are similar in tissue morphology, pigment, etc., so this represents “limited knowledge” from the perspective of a DL model; there is a need for larger case series, with multicenter validation
Liu et al. (2022) [13]	Considerable sensitivity, accuracy, and specificity for predicting GEP in UM from digital cytopathology images	Small or thin UM tumors may not be suitable for FNABs, which are necessary to produce cytopathology slides. The DL algorithm was trained and validated using cellular aspirates stained with H&E. To ensure compatibility with specimens prepared using Papanicolaou staining, the algorithm requires further adaptation. Additionally, the study population lacked substantial diversity, highlighting the need for prospective validation to confirm the model’s generalizability
Akram et al. (2024) [14]	Predicting UM molecular subclasses by CNN using routine H&E-stained slides with promising results	There was a limited dataset size, which requires further training and validation on external cohorts
Liu et al. (2020) [15]	Predicting GEP in UM using DL models based on H&E-stained slides from FNAB	FNABs can be technically challenging and susceptible to sampling errors. There was a small patient sample size. There is a need to train an algorithm with a larger database and actual survival data
Reggiani et al. (2024) [17]	Developing ML models to identify molecular classes of UM based on different gene alterations	Technical variations in sequencing experiments can obscure or imitate biological differences, potentially leading to model overfitting. This may hinder the extraction of meaningful biological features and limit the model’s ability to perform accurate classification when applied to a different dataset. Furthermore, it remains uncertain whether molecular classification approaches would retain their efficiency when applied to larger datasets, necessitating further investigation
Xie et al. (2020) [18]	Employing bioinformatic analysis to identify key genes in metastatic UM	All data were obtained from the Gene Expression Omnibus (GEO) database rather than directly from tissues of UM patients. All conclusions were based on bioinformatic analysis; therefore, caution must be exercised in interpreting the results. The dataset consisted of 39 primary tumor samples and 2 metastatic samples, so there may be a possible bias due to the imbalance between groups
Hou et al. (2021) [19]	Identifying DNA methylation signatures in UM, which may be of prognostic relevance, stratifying patients into two risk groups, with significantly different prognosis in terms of survival.	The discovery of the DNA methylation signature was based on in silico analysis. Independent validation of the ML model was challenged by the lack of an available patient cohort with clinical information

4.3. AI in Prognostic Prediction

The possibility of applying AI models to identify prognostic factors in patients affected by UM has been investigated in different studies during recent years. In 2022, Bassi et al. analyzed the possibility of predicting treatment outcomes in patients with UM. In particular, they highlighted the role of AI models in weighing the significance of biomarkers, such as lactate dehydrogenase (LDH), against other prognostic variables with respect to treatment outcomes [20].

Lever et al. analyzed hepatic metastases from UM by using ML tools, showing that substantial binary risk stratification can be performed, based on (i) clinical and laboratory parameters, (ii) measures of quantitative overall hepatic tumor burden, and (iii) radiomic parameters [21]. However, the combination of two or all three domains failed to improve prognostic stratification. They also identified highly relevant parameters, including LDH, platelet count, aspartate aminotransferase (AST), and the metastasis-free interval at the first diagnosis of metastatic disease, as predictors for time-to-treatment failure and overall survival (OS). The risk stratification models based on this ML algorithm identified a comparable and independent prognostic value of clinical, radiological, and radiomic parameters in UM patients with hepatic metastases.

In recent years, there has been growing interest in utilizing AI to identify factors predictive of metastatic disease. In a study aimed at predicting survival in patients with uveal melanoma (UM), Chen et al. employed random forest, a classical machine learning (ML) method, to develop two predictive models for estimating the risk of death and metastasis within two years [22]. These models were trained using demographic characteristics, ophthalmic examination variables (notably, intraocular pressure), and tumor-specific features, including largest basal diameter, thickness, pigmentation, location, macroscopic appearance, optic nerve and ciliary body involvement, subretinal fluid, intraocular hemorrhage, extraocular extension, TNM stage, and initial treatment. The study identified largest basal diameter, thickness, tumor size, intraocular pressure, and initial treatment as the most significant parameters. For death prediction, the model achieved an AUC of 0.883 and an accuracy of 76.9%, while for metastasis prediction, it achieved an AUC of 0.846 and an accuracy of 74.9%.

Similarly, Donizy et al. utilized classical ML models to identify predictors of metastasis and survival in a cohort of 164 patients with UM who underwent ocular enucleation without prior treatment [23]. Three models were constructed to identify predictors of overall survival (OS) and progression-free survival (PFS). Across all three models, the following features emerged as significant predictors of OS: age, ciliary body infiltration, mitotic rate per 1 mm², BAP1 status, and nucleoli size. For PFS, the most significant predictors included tumor-infiltrating lymphocyte and macrophage densities, largest basal diameter, nucleoli size, and BAP1 status. However, a potential limitation of this study was the inclusion of only patients who underwent ocular enucleation. This may introduce selection and outcome biases, as current UM management more commonly involves local therapies such as plaque brachytherapy or proton beam radiation rather than enucleation.

In contrast, Luo et al. developed a machine learning model to predict the 4-year risk of death and metastasis in patients with UM treated with iodine-125 plaque brachytherapy [24]. This model incorporated demographic data (age and sex), ocular features (laterality, corrected visual acuity, and intraocular pressure), and tumor characteristics (presence of subretinal fluid, optic nerve and ciliary body involvement, vitreous hemorrhage, tumor thickness, tumor shape, and basal diameter) derived from multimodal imaging. Notably, the model’s performance improved significantly when data from three follow-up visits were included, rather than data from a single visit. With this additional postoperative information, the model achieved an accuracy of 83% for predicting death and 79.5% for predicting metastasis. These findings underscore the importance of incorporating longitudinal follow-up data to enhance the accuracy of predictive models for UM-related outcomes.

In a recent retrospective study, Wu et al. investigated potential risk factors for metastatic disease in patients with UM using ML models [25]. The study identified higher tumor staging, specific histologic features (mixed epithelioid and spindle cells, as well as epithelioid cell melanoma), primary tumor location in the ciliary body, surgical treatment, and systemic therapy as significant risk factors for distant metastasis (p-value < 0.05). The models achieved an accuracy of 78.8%. One of the study’s key findings was the identification of the ciliary body as the tumor location associated with the highest metastatic risk, potentially due to the distinct vascular supply of the anterior chamber and iris compared to the choroid. Another critical insight from this study was the importance of early diagnosis and treatment in preventing distant metastasis. Radiation therapy, in particular, was highlighted as playing a pivotal role in reducing metastatic risk by controlling local tumor growth, inhibiting angiogenesis, inducing apoptosis, and stimulating immune system activation.

Damato et al. utilized a neural network to estimate survival probabilities following surgical treatment for UM and stratified risks based on various clinico-pathological features. The neural network generated survival curves comparable to Kaplan–Meier estimates, indicating that predicted survival closely aligned with observed outcomes [26].

Another significant challenge in UM research involves the potential malignant transformation of a choroidal nevus into melanoma. The accurate prediction of this transformation carries substantial implications for treatment strategies, ocular morbidity, prognosis, and mortality. Tailor et al. addressed this issue by developing an ML model designed to predict the malignant transformation of choroidal nevi. The model, based on tabular multimodal imaging data from a multicenter study with internal and external validation, achieved a sensitivity of 86.9% and a specificity of 98.4% [27]. Key predictive factors for malignant transformation included lesion thickness, lesion diameter, and tumor shape (particularly, dome-shaped lesions). This area remains a critical focus in ophthalmic pathology, requiring further investigation due to the current lack of definitive criteria for distinguishing between a choroidal nevus and a small UM.

The primary AI applications, along with their respective advantages and limitations, in predicting UM prognosis and metastatic risk are summarized in Table 3. AI shows promise in predicting metastatic risk and overall survival by leveraging clinical, imaging, and histological data. Models have demonstrated good accuracy and can support early intervention strategies. However, many studies are retrospective with limited diversity, potential selection bias, and insufficient external validation, reducing generalizability.

Table 3. AI applications to predict UM prognosis and metastatic risk.

Study	Advantages	Limitations
Bassi et al. (2022) [20]	Large amount of information regarding significant biomarkers	There was no mention of the reliability of the biomarker sample collection process. There was a large degree of heterogeneity
Lever et al. (2024) [21]	Using ML algorithms to stratify prognosis of UM patients with liver metastasis, identifying clinical, laboratory, and radiomic parameters to predict overall survival	This was a retrospective analysis, with the need for external validation. Combining parameters does not increase the risk stratification accuracy. Training and test groups were not ideally balanced with respect to gender and global outcomes, leading to a smaller proportion of high-risk patients in the test group and, more importantly, a different median risk score
Chen et al. (2022) [22]	Identifying clinico-pathological features which may be predictive of death and metastasis in UM; high sensitivity and specificity	This retrospective study was conducted only on a Chinese population
Donizy et al. (2022) [23]	Three ML models which identified predictive features for OS and PFS in UM	Possible bias: only patients who underwent ocular enucleation were included in the study
Luo et al. (2022) [24]	Prediction of risk of death and metastasis in patients with UM who have undergone plaque brachytherapy	The possibility of metastatic disease at diagnosis may affect the model’s predictive value for metastasis. Due to the retrospective nature of the study, the follow-up interval after surgery was not fixed.
Wu et al. (2024) [25]	Developing ML models to identify risk factors for metastatic UM; highlighting differing metastatic potential between ciliary body and choroid melanoma	This was a retrospective analysis, with potential information and selection bias difficulties in establishing causal relationships. There was a lack of external validation with data from different hospitals. The performance of ML models may vary across patients from different regions and hospitals
Damato et al. (2008) [26]	Using neural network to estimate survival probability after treatment of UM, generating curves which were similar to Kaplan–Meier	Some combinations of risk factors were rare, so survival could be modeled only imprecisely. There is a need for external validation. Data were obtained from patients treated by enucleation or local resection. There is a need for further studies to determine whether survival predictions would be accurate in patients undergoing non-surgical treatment
Tailor et al. (2024) [27]	First study developing ML models to predict malignant transformation of choroidal nevi into melanoma; large multicenter cohort, use of multiple imaging modalities to comprehensively phenotype nevi, and external validation demonstrating model robustness	The retrospective design of the studies, which rely on existing records, introduces inherent limitations, including a lack of tissue confirmation of the disease in most cases. Both datasets predominantly represent a White population, limiting the generalizability of findings to more diverse populations. The reliance on clinician annotations and tabular data, rather than the direct extraction of raw imaging features, may reduce the potential for uncovering novel predictive markers. Furthermore, the imbalanced nature of the classification problem, due to the relatively rare occurrence of transformation events, poses challenges for model training and evaluation. Without further validation and evidence of clinical utility, there is a risk of overstating the practical value of these approaches in guiding clinical integration

5. Conclusions

Various types of AI algorithms were evaluated in the included studies. Compared to ML, DL algorithms require larger datasets for training to achieve optimal performance. Consequently, when training datasets are limited or of suboptimal quality (e.g., lack of heterogeneity), the performance of these algorithms may be adversely affected.

In recent years, AI has increasingly been recognized as a valuable tool for the diagnosis and management of various tumors, including central nervous system tumors and head and neck cancers [28,29].

The application of AI in the histopathological and molecular diagnosis of UM, similar to its use in other healthcare domains, presents several technical, regulatory, and clinical challenges. From a technical perspective, AI models, particularly those based on DL, require extensive, well-annotated datasets, which are difficult to standardize across institutions, especially for rare diseases like UM. Additionally, DL models are often perceived as “black boxes”, making their decision-making processes difficult to interpret, which can be a concern for clinicians who require transparency to trust AI-driven diagnoses. Furthermore, the high computational requirements of complex AI models may limit their feasibility in smaller healthcare settings.

Regulatory challenges include ensuring data privacy, maintaining security, and adhering to established standards. Certifying AI models for clinical use is particularly challenging for evolving systems, and liability concerns related to potential AI misdiagnoses remain unresolved. Clinically, one of the main obstacles is gaining the trust of healthcare professionals, who may hesitate to rely on AI tools if the underlying algorithms are not easily interpretable. Moreover, AI models need to integrate seamlessly into existing clinical workflows to avoid disruptions. Ensuring that these systems generalize well and perform effectively across diverse patient populations is essential for their widespread adoption in clinical practice.

Further research is necessary to explore AI applications in UM. Longitudinal studies assessing the long-term impact of AI on patient outcomes are particularly important. It is also crucial to evaluate how AI-assisted diagnoses influence treatment decisions, disease progression, and patient outcomes. Including diverse patient populations in studies is another critical factor to ensure AI models are trained on representative datasets, thereby reducing biases and improving their generalizability. Notably, AI can facilitate collaboration among physicians from various specialties (e.g., oncologists, radiologists, pathologists, and surgeons), fostering the development of integrated and comprehensive care strategies.

Finally, the ethical and social implications of AI adoption in healthcare cannot be overlooked. Research must address concerns related to patient privacy and the potential for algorithmic bias, as well as investigate how AI integration might influence equity and access to care across different populations. It is important to note that AI algorithms, particularly those based on deep learning (DL), have been increasingly applied in recent years for the detection of uveal melanoma (UM) and the differentiation of benign tumors, such as choroidal nevi, from UM. The applications of AI in UM extend beyond histopathological diagnosis to include molecular analysis, such as gene expression profiling (GEP) and methylation patterns, which are crucial for predicting patient prognosis.

In summary, while AI holds great potential to revolutionize the diagnosis and prognosis of UM, coordinated efforts are essential to overcome current challenges related to data standardization, model interpretability, and regulatory compliance. By addressing these critical issues and emphasizing future research that incorporates population diversity and ethical considerations, we can facilitate the seamless integration of AI into clinical workflows. Ultimately, this will enhance early detection, optimize treatment strategies, and improve patient outcomes in UM management.