Every year, the eye care industry generates a vast number of ophthalmic images, including fundus photos, OCT scans, slit lamp microscope images, and more. The effectiveness of patients’ treatment relies heavily on the precise interpretation of these images by optometrists.
However, as the volume of medical images continues to rise, so does the risk of medical errors. Today, there exists an opportunity to enhance the speed and accuracy of medical image interpretation while reducing the risk of medical errors through the application of artificial intelligence (AI).
AI makes it possible for machines to learn from experience, adjust to new inputs and perform human-like tasks. Many AI examples that are more common today – from apps on your mobile phone to self-driving cars – rely heavily on deep learning and natural language processing.
Using these technologies, computers can be trained to complete specific tasks by processing huge amounts of data and recognising patterns in the data. It does not mean that AI will analyse medical images instead of eye care specialists: it simply means that AI can help in triaging patients correctly and advising on the interpretation.
The use of AI in interpreting ophthalmic images began with the analysis of fundus photos for diabetic patient screening. In some parts of the world, fundus AI systems have been designed for non-eye care settings, such as for primary care physicians, who might provide screening services for diabetic patients.
In traditional optometry practices, these systems could also be used to facilitate screening and diagnosis. However, fundus photography is often complemented by optical coherence tomography (OCT) imaging in a modern eye care practice as a more informative diagnostic modality.
OCT stands out as one of the most effective diagnostic techniques and yet is one of the most challenging to become an expert in. It requires years of learning and practice for optometrists to master. As a result, a variety of AI tools have emerged to aid optometrists with interpreting OCT scans.
These tools fit various purposes and have different features and limitations. However, they are united by the benefits they bring to the optometry practice.
Potential benefits of AI-powered tools for OCT analysis
Why should optometrists consider using AI-powered tools for OCT scan analysis? There are number of potential benefits:
- Workflow efficiencies with the speed and accuracy for interpreting OCT scans therefore helping to build optometrist confidence.
- Detection of early and rare pathologies/biomarkers, which could mean better standard of care for the patients.
- Reduced risk of missing pathology in practice and, therefore, lower clinical risk and governance.
- Expert ‘second opinion’ from AI regarding complex OCT scans could assist the optometrist with diagnostic support and referral refinement.
- Improved standard of care for patients, which increases patient satisfaction and loyalty.
- Referral for efficient automated triage: creating a refinement of a referral process.
- Reduction in the number of false-positive patients referred to the NHS.
- More effective co-management: optometrists could be sure that they send patients to ophthalmologists only when there are conditions that need referral.
Potential limitations of AI-powered tools for OCT analysis
- Usually, AI analysis is based on one image modality and does not take into account other clinical inputs
- Data quality impacts AI analysis results. Better quality of OCT scans will mean better response from the AI algorithm
- There is a gap between lab AI models and clinical use. Research models have high accuracy but may be developed on precisely selected high-quality data.
- In many cases, AI systems should be used as a recommendation system, autonomous AI systems have very narrow use.
- Regulations can limit the utilisation of certain AI systems.
AI-powered tools for OCT analysis available on the market
AI tools for OCT analysis can prove invaluable for informed clinical decision-making. This not only reduces the screening workload but also affords more time for direct patient care.
The shift facilitated by these AI tools transforms the role of eye care professionals from primarily screening and diagnosing to a focus on treatment and, ultimately, the prevention of vision loss.
There are an increasing number of tools becoming available in the UK, for example, with many more in development. Each tool has its own specific characteristics; the range of eye conditions tools can detect, diagnose and grade conditions can differ. Other features include the potential for referral support and tracking over time.
A key point when considering the implementation of OCT in your practice is compatibility of the AI tool with your existing OCT equipment and practice software.
Glaucoma risk detection
AI accuracy
There are different ways to estimate the accuracy of AI models, for example Dice and F1 scores, which are made up of two key
elements:
- Precision: the number of times the machine learning model makes a correct prediction.
- Recall: the number of predictions in a dataset that the machine learning model identifies correctly.
Both these formulas work with false positive (FP) and false negative (FN) errors. The false positive (FP) score measures how many times the AI model detected a pathology/biomarker when there was none. The false negative (FN) score measures how many times the AI model did not detect the pathology/biomarker when there was one.
The Dice score is used to quantify the performance of image segmentation methods and varies between zero and one with one meaning perfect correlation. It is impossible to use a single number when quantifying the performance of the AI algorithm because the model comprises several algorithms and the score would be different for each pathology and a biomarker.
There are an increasing number of studies supporting the use of AI tools for OCT interpretation. A recent systematic review and meta-analysis evaluating AI performance for detection of diabetic macula oedema found a sensitivity of 95.9% and specificity of 97.9%.1
Similarly, another review and meta-analysis also considering AI analysis of diabetic retinopathy reported a pooled sensitivity and specificity of 96% and 99.3% respectively.2 A review of studies focusing on conditions such as pathologic myopia has also shown favourable outcomes.3
Misconceptions regarding AI for OCT scan analysis
There is already a variety of AI-powered tools for OCT scan analysis available, and while some optometrists have started using them in their clinical practice others may still have preconceived ideas about AI in healthcare. For instance, professionals may be worried that AI will replace them if they start using it. A few common misconceptions about AI are discussed below.
First and foremost, AI is unlikely to replace the expertise of human specialists. The evolution of AI is constant and unstoppable. Yet for a conclusive diagnosis, optometrists must consider various factors, including clinical history, examination results, additional diagnostic findings, comorbidities and more. In every instance, the definitive decision-making should remain with the eye care professional.
Secondly, AI may not always be designed for prediction but may be for in-depth image analysis. Creating AI tools that predict outcomes based on medical imaging can be challenging. Numerous factors, unknown even to clinicians, and the highly individualised nature of disease progression make such predictions complex.
Although there are numerous research projects in this realm, the transition from research to practical applications remains a significant challenge.
Third, using AI is not particularly complex. AI can seem a difficult tool to use in practice but optometrists can start using it now. AI does not require extraordinary skills, it requires high-quality data (such as Dicom files from your OCT device), a critical mind, and some time to get used to a new technology.
Pathologies biomarkers
The Future of AI in OCT Scan Interpretation
AI is here to stay and will likely become more accurate and be able to work with a larger range of biomarkers. Recently there have been reports about how AI algorithms will be able to detect early biomarkers for neurological diseases such as Parkinson’s disease, Alzheimer’s, and other debilitating diseases.
These AI models may eventually become available in practice. Not only will the advances in AI make optometrists’ lives easier but they will provide the quality experience that can improve the standard of care and help build long-term patient loyalty.
Although AI has lots of potential to improve healthcare, it does come with limitations and ethical considerations. For example, what if the interpretation of symptoms generates an inaccurate result? As stated earlier in the article, the decision-making must remain with the optometrist; they need to review the data to confirm the decision.
Patient data privacy is important to make sure that the technology complies with the current regulations. With the possibility of huge amounts of data coming in from various sources, organisations must also invest in the technology needed to keep it secure.
The popularity and widespread availability of AI tools in the eye care industry is inevitable. While currently there are only around six AI tools for the OCT platform, their number will grow from year to year. Accessibility to such tools in terms of price and user-friendly design will make AI systems commonplace for optometry practices in the near future.
Optometrists are already embracing AI tools, such as ChatGPT, so OCT analysis with AI will not be a difficult transition in just a couple of years’ time.
- Maria Znamenska, MD, PhD, is an associate professor of ophthalmology and chief medical officer at Altris Inc. She is an ophthalmologist with over 16 years of experience. She has been working with OCT since 2005 and leading a medical stream of R&D of the US health-tech company Altris Inc. since 2017. Her medical team launched Altris Education OCT and OCT-angiography training courses for ophthalmologists in Eastern Europe and CIS region in the beginning of 2020.
References
- Lam C, Wong YL, Tang Z, Hu X, Nguyen TX, Yang D, Zhang S, Ding J, Szeto SKH, Ran AR, Cheung CY. Performance of Artificial Intelligence in Detecting Diabetic Macular Edema From Fundus Photography and Optical Coherence Tomography Images: A Systematic Review and Meta-analysis. Diabetes Care. 2024 Feb 1;47(2):304-319. doi: 10.2337/dc23-0993. PMID: 38241500.
- Li HY, Wang DX, Dong L, Wei WB. Deep learning algorithms for detection of diabetic macular edema in OCT images: A systematic review and meta-analysis. Eur J Ophthalmol. 2023 Jan;33(1):278-290. doi: 10.1177/11206721221094786. Epub 2022 Apr 27. PMID: 35473414.
- Zhang Y, Li Y, Liu J, Wang J, Li H, Zhang J, Yu X. Performances of artificial intelligence in detecting pathologic myopia: a systematic review and meta-analysis. Eye (Lond). 2023 Dec;37(17):3565-3573. doi: 10.1038/s41433-023-02551-7. Epub 2023 Apr 28. Erratum in: Eye (Lond). 2023 Dec 13;: PMID: 37117783; PMCID: PMC10141825.
