As artificial intelligence (AI) continues to reshape the landscape of retinal disease, retinopathy of prematurity (ROP) sits at a crucial intersection of technological innovation, clinical nuance, and global health demand. J. Peter Campbell, MD, MPH, offered a forward-looking overview of how AI is being developed, validated, and deployed to strengthen ROP care in a presentation at the FLORetina 2025 meeting and International Congress on OCT Angiography, En Face OCT and Advances in OCT (ICOOR), held from 4 to 7 December in Florence, Italy.¹ Campbell is the Edwin and Josephine Knowles Endowed Professor of Ophthalmology and professor, Casey Eye Institute, Oregon Health & Science University in Portland, Oregon, USA.

The Eye Care Network spoke with Campbell to delve deeper into the realities of bringing these tools from research to clinical practice—from regulatory hurdles and performance validation to issues of bias, generalisability, and implementation in diverse health care settings. The Q&A that follows highlights his perspectives on the most promising directions for AI in ROP and the challenges that must be addressed as the field moves ahead.

Note: Transcript edited lightly for clarity and length.

What specific AI models have shown promise in ROP detection and grading?

J. Peter Campbell, MD, MPH: As background, it is helpful to think through the various ways AI can be helpful in the context of retinal diseases in general because the applications are myriad. This means that “AI in ROP” can mean any number of actual products. In the US FDA context, image-based AI algorithms are generally considered Software as Medical Devices (SaMD), and regulated by the FDA depending on the indication for use (IFU). To oversimplify, AI algorithms can be used for screening (refer or not) or diagnosis (any part of the International Classification of Retinopathy of Prematurity—classification of zone, stage or plus), and can be used assistively (aiding a clinician) or autonomously (no clinician in the loop) for referral decisions.

There is a huge translational gap between papers published showing effectiveness of an AI algorithm, evidence of effectiveness in external real-world datasets, and SaMDs that have traversed the regulatory pathway to be useful clinically. To date, no AI algorithms have been approved for ROP in the US or Europe. I’ll try to answer with these concepts in mind.

There have been numerous academic papers describing AI algorithms that have shown promise for ROP screening or diagnosis. The most common approaches have been image-based algorithms that identify the classic ROP “stage” lesion in images, diagnose plus disease, or assign a 1-9 vascular severity score to images. There are 2, to my knowledge, that are navigating the regulatory pathway towards clinical use.

First, Remidio has developed an on-device AI algorithm that highlights the appearance of the vascular-avascular border and any stage in ROP images. It does this with a heatmap to be used by the clinician to better detect the presence of ROP in images.

Second, Siloam Vision is commercializing the iROP-DL algorithm that has 2 potential IFUs, autonomous screening, and assistive diagnosis. The assistive version of the product outputs a 1-9 vascular severity scale from input Retcam images. This output has been shown to improve the clinical diagnosis of plus disease among ROP clinicians. Since interobserver variability in plus disease is a well-known problem, this may improve the precision and accuracy of diagnosis in the future. The autonomous version outputs a “refer, or repeat in 1 week” decision to the photographer. If “more than mild ROP” is detected in a set of images, then the images are forwarded for telemedical review. If not, then the baby is scheduled for repeat evaluation in 1 week. This has demonstrated 100% sensitivity for detection of treatment-requiring ROP both in 2 large datasets of more than 5000 consecutive examinations in the US and India.

It is important also to note that AI algorithms that work on 1 camera may not work on images from another camera, and so effectiveness must be evaluated in that context. For the iROP-DL algorithm, we have published data demonstrating the performance for assistive diagnosis on the Retcam, and autonomous screening on the Retcam, Forus, and Icon cameras.

How do you validate algorithm performance against expert consensus?

Campbell: This is an important question, especially for ROP, as interobserver variability is a problem, which means that comparing to a single label is problematic. The FDA has very clear guidance for reference standards, with the gold standard being a standalone reading center with clear metrics for the methodology of assigning a diagnosis. For the iROP-DL algorithm, we used a 5 expert reference standard to plus disease.

What are the barriers to implementing AI in low-resource settings?

Campbell: There are several. The biggest barrier is the cost of current ROP cameras. The ideal would be a camera that could provide digital images with widefield or ultra-widefield visualization and be placed in every neonatal unit in the world at a reasonable cost. Current ROP cameras are likely too expensive to make this a reality, but several groups are trying to develop lower cost cameras.

The next barrier will be validation of AI on the target low-cost device, and determining the precise indication(s) for use for AI. Are we wanting to trust AI to do the screening without a physician (autonomous use)? Are we wanting to use it to standardize clinical diagnosis to improve the sensitivity and specificity of clinical diagnosis (assistive use), or some combination thereof?

Finally, one of the challenges with AI implementation in general is the infrastructure that enables integration of AI into a workflow that is easy to adopt and use by all of the key stakeholders. That said, this is the direction we need to go—ROP blindness is the most solvable problem in global ophthalmology.

How do you address concerns about bias and generalisability in AI-driven diagnostics?

Campbell: Two huge topics. Briefly, AI algorithms can be biased in multiple ways. Most commonly, we think of issues related to differences in performance by racial/ethnic groups, but this is only one of many potential biases that can be trained into AI algorithms. There are well established best practices to minimize bias, including the use of reference standard diagnostic labels and the evaluation of performance in multiple subgroups to ensure equitable performance across populations.

It is also important to remember that human clinicians are biased as well. For example, it has been shown that clinician knowledge of gestational age impacts image-based diagnosis of plus disease, which should be independent, so as we think through implementation of these technologies it is critical to remember that the status quo is imperfect.

In terms of generalizability, AI algorithms that work well in one context may not work in a context that varies from any of several variables including image quality, technical specifications of various cameras, illumination differences, use of various segmentation tools, demographic population differences, etc. This is why the regulatory bodies, such as the FDA, are rightly focused on ensuring that the data used for regulatory approval reflect the type of data that an algorithm will see in its intended use in the intended population.

Are there global or regional barriers that are specific to ROP?

Campbell: The epidemiology of ROP is such that the largest potential need is in low- and middle-income countries, where there are an order of magnitude higher number of premature births, and an order of magnitude lower number of ophthalmologists, making AI-enabled ROP telemedicine an ideal solution especially since the population is captive (in neonatal units). There are successful models of ROP telemedicine working (without AI) so it is really important to highlight that the lack of “AI” is not the main issue—it is the lack of infrastructure and/or the capital and human resources to initiate and scale ROP telemedicine itself. AI can add value in a number of ways, but the key barrier is the underlying infrastructure.

What future applications do you envision beyond screening?

Campbell: “Screening” has a very particular definition from the regulatory perspective. That is, a “screening” test is one that happens prior to definitive clinical diagnosis. In the context of ROP, this means a “refer or not” decision similar to what we are seeing with diabetic retinopathy.

We have seen autonomous diabetic retinopathy companies successfully navigate the regulatory pathway and enter the market, for example. It will be interesting to see if, and in which contexts, autonomous ROP screening might be viable. It is one thing for it to work, theoretically. It is another thing to have societal and commercial adoption given medicolegal risks, among other things. In that sense, the use of AI to aid “diagnosis” may be higher impact since the decision making ultimately remains with the clinician responsible for the care of a particular baby.

In addition, there has been work utilizing AI-based biomarkers in the context of risk modeling to reduce the number of exams needed for low-risk babies, and even potentially discontinue screening. We published data from the US, India, Mongolia, and Nepal demonstrating that in theory a single exam could rule out clinically significant ROP in low risk babies in all populations.

Finally, there is emerging evidence that images of retinas can contain information relevant not only to the diagnosis of ROP, for example, but also of systemic diseases. It will be interesting to see if and where that potential capability of AI adds clinical value in this context.

What’s the most valuable perspective you hope to gain from experts in both adult and paediatric retina?

Campbell: I find that meetings like this provide intangible value just by bringing together an incredibly diverse group of retina specialists from a variety of geographic and practice settings. Especially for rare diseases, this allows for collaboration and opportunities to advance the field which are not easy to do at a single center. For ROP, we all have a lot to learn from each other. As an adult and paediatric retina specialist, it is also exciting to see the advances in retinal imaging that are often introduced to adult patients first and then eventually make their way to paediatrics, OCT being one example.

Are there any emerging therapies you expect to see discussed in both an ROP context and within the broader FLORetina programme?

Campbell: As with ophthalmic imaging, changes in ROP therapeutics often follow changes in adult retinal therapeutics. However, there are a few potential therapeutic targets and approaches that may soon be in clinical trials in ROP that are unique to this population. It will be exciting to see whether the investigators and companies involved in these approaches are ready to present them at this year’s meeting.

J. Peter Campbell, MD, MPH
E: [email protected]
Campbell is the Edwin and Josephine Knowles Endowed Professor of Ophthalmology and professor, Casey Eye Institute, Oregon Health & Science University in Portland. Campbell notes he is affiliated with Siloam Vision that is developing some of these technologies.

REFERENCE

1. Campbell JP. AI on ROP. Presented at: FLORetina 2025; December 4-7, 2025; Florence, Italy.