Artificial Intelligence in Ocular Oncology and Pathology: A Call for Contributions

Artificial intelligence is arriving in medicine faster than any of us can keep up with. In the past 2 years, the largest technology laboratories have released general-purpose models that read whole pathology slides [1], predict the pathogenicity of nearly every human missense variant [2], and reason through clinical cases at the level of specialists in ophthalmology [3]. Ophthalmology has been an early beneficiary: autonomous AI for diabetic retinopathy has been authorized for primary care since 2018, the first autonomous diagnostic AI cleared in any field [4]. Ocular oncology and ophthalmic pathology are harder tests, and not mainly because the tumors are rare. The deeper reason is that the field is small: a close community with real depth of expertise, but few hands; its data held in a handful of centers; and in much of the world, no ocular pathologist or ocular oncologist within reach. This editorial takes stock of what artificial intelligence has accomplished in ocular oncology and pathology, how far the newest tools remain from routine use here, and where the most useful contributions now lie.

The progress is real, but its scope remains narrow. Computer vision is the mature layer: convolutional networks match or exceed clinicians on discrete tasks, from retinoblastoma screening to smartphone leukocoria detection to eyelid and choroidal tumors [5–9], and vision transformers now read uveal melanoma subclass and ocular-surface neoplasia directly from routine images [10, 11]. Genomics is a second bright spot: AlphaMissense, which scores nearly every human missense variant [2], has been carried into uveal melanoma to resolve variants of uncertain significance in BAP1 and its companion genes, holding up even against experimental gold standards [12, 13]. Retinal foundation models, such as RETFound, are built on retinal images and then fine-tuned to tell choroidal melanoma from nevus [14, 15]. Yet, these successes share a common bottleneck: they are largely confined to single-center data, rarely reach external validation, and have yet to translate into the multimodal, prospective trials required for routine practice.

While these specialized models represent progress, broader computational breakthroughs emerging across general medicine remain entirely unrealized in our field. The frontier laboratories continue to ship models, yet their reach into ocular oncology and pathology remains limited. The whole-slide foundation models now transforming computational pathology, Prov-GigaPath [1], the multiplex-generating GigaTIME [16], and general-purpose models like UNI and Virchow [17, 18] have not, to our knowledge, been reported on a single ocular specimen. Large language models perform at a specialist level in general ophthalmology yet appear in only a handful of ocular oncology studies, none of which touch ophthalmic pathology [19–21], and agentic systems, now spreading through oncology and pathology, are absent here altogether [22, 23]. The reason is structural. These models are built the way large fields are, on broad consortia and shared archives [1, 16]; ocular oncology, smaller and more scattered, has not assembled its data in the same way. Even its central genomic reference, a single cohort of eighty uveal melanomas [24], reflects how little of the field’s material has ever been pooled, not a hard limit on the cases that exist.

Bridging these translational and technological gaps is precisely why the journal’s new artificial intelligence section was created. The stakes are not purely academic: built well, these systems may let a clinician who is not an ocular pathologist, in a place with no ocular oncologist within reach, act on a finding that would otherwise wait weeks or months, and in these tumors, time lost is vision and sometimes life. The section cannot organize the field on its own, but it can gather this work, hold it to a standard, and give it a home. The call for contributions is open.

Figure 1 maps a non-exhaustive framework of the approaches and open problems we hope to see addressed. While these core domains represent immediate gaps, we welcome work at every stage. We are especially interested in studies that carry a method a step further. This includes external validation across different cohorts that shows a promising model holds up beyond the center that built it, work that fine-tunes general pathology foundation models specifically for ophthalmic tissue, and research that extends variant and multi-omic prediction across ocular tumors, somatic and germline. A well-designed single-center study is a valuable contribution, and our aim is simply to help such work grow into tools that others can trust and use.

Fig. 1.

Artificial intelligence in ocular oncology and pathology: a representative framework of method classes and open problems invited by the section.

Our interest reaches beyond images and sequences to work closer to the clinic. Possible directions include workflows for plaque brachytherapy planning, automated fine-needle aspiration biopsy guidance, and prognostic tools that leverage retrospective registry data to aid clinical decision-making.

We are equally interested in the language and agentic systems that help organize data on ocular tumors, whether for drafting pathology reports or for synthesizing imaging, histology, and molecular findings into multimodal clinicopathological correlations. Furthermore, we welcome the datasets and collaborations on which all of this depends: multi-center consortia, open and curated registries, and federated learning approaches that let institutions train models together while their slides and data remain within secure local repositories [25]. Any rigorous application of artificial intelligence that organizes data in ocular oncology and pathology or applies a frontier model to an eye tumor is welcome in these pages.

Acknowledgment

During preparation of this manuscript, the author used generative AI tools to assist with language refinement. All content was reviewed and edited by the author, who takes full responsibility for the accuracy, integrity, and originality of the work.

Conflict of Interest Statement

Mak Djulbegovic was a member of the journal’s Editorial Board at the time of submission.

Funding Sources

This study was not supported by any sponsor or funder.

Author Contributions

M.B.D. conceived, drafted, critically revised, and approved the final manuscript.

Funding Statement

This study was not supported by any sponsor or funder.