Ophthalmic education, Optical Dispensing

How AI is transforming spectacle lens design

AI Lens design

 

In optics, AI has not only answered questions sought out by researchers but has identified unexpected correlations between different eye parameters. What are the implications for future lens design?

Artificial intelligence (AI) is easily one of the biggest buzzwords in the optical industry of late. The processes are significantly increasing the pace of development in lens design, however, a full discussion of AI is well beyond the scope of such a short article so we will cover the broad strokes here.

Grant Hannaford. Image: AAOO.

“AI applications for lens design are primarily used to process bulk data to address a specific question posed by a design team.”

There is a common misconception that AI is able to answer any question rapidly, but like any of us, the answer generated is only as good as the information available to it. In many cases, the information used is drawn from the internet, so the answers are subject to the general noise and misinformation present online. An easy example of this is to ask for definitions of simple optical conditions, such as hypermetropia. The answers, while close, miss the mark by just enough to show the answers aren’t being completely drawn from clinical sources. In the same way, for lens design we require a good database for generation of results (more on that later).

The overall areas in which AI is currently providing advances in lens design, generally, are:

• Best form optimisation, where the relationship between Rx and design are optimised on a per case basis.

• Position of wear optimisation, using the full suite of positioning and power interactions to develop compensate designs.

• Progressive addition lens (PAL) design, similar to best form where add, Rx and design interactions are optimised.

• Behavioural modelling and biometric modelling, in which the specific biometric and behavioural requirements of a patient are used to modify lens designs and individualise layouts.

All of these areas are providing significant patient improvements and are the subject of a presentation available via the Academy of Advanced Ophthalmic Optics.

AI applications for lens design are primarily used to process bulk data to address a specific question posed by a design team. This facilitates searching large data sets with complex interactions for elements that may influence the success or failure of a design. In a soon-to-be-published study, a series of environmental and lifestyle questions are related to biometric factors in the eye, with the total number of potential interactions between factors for one candidate alone in excess of 40,000. Across a cohort of individuals it is not possible to work through these data sets reliably, so AI provides iterative processes that can not only answer questions developed by the researcher, but also have been shown to identify unexpected relationships in the data sets as well.

It would seem ideal to have every single individual set of Rx, behavioural and biometric conditions included in the data, however, a sufficiently strong outlier can skew the results significantly. For example, a model can be built relating refractive error to axial length which can inform personalised best form designs. If an eye deviates significantly from the model, for example very short but highly myopic, the determined axial length for that power will shift sufficiently. In a research sense, this informs our understanding of biometry, but for a practical applied design sense it is undesirable. The outlier can change the model being used, making the model inaccurate for the bulk of the other patients using the lens. In other words, we need to ‘clean’ the input data to ensure that the information being used to generate the designs is applicable to the largest set of the population possible. This is also why it is not possible to allow an AI process to continually draw in new data ‘unsupervised’, as the models it generates can become skewed. Typically, teams will engage in a quality assurance program to assess the output of the AI data analysis prior to inclusion in a lens design, with data that is leading to a skewed output being diverted for alternative analysis.

Iterative processes are not new in data management, but the ability of AI to isolate relationships previously hidden to researchers is allowing the pace of development for lens design to increase. The opportunities this affords our industry are significant to say the least. 

NOTE: For more detail on the processes themselves, Matlab has an excellent tutorial covering the basics, contact AAOO for the link.

About the author: Grant Hannaford is a qualified lens designer, has completed an MSc (optometry) and is undertaking Doctoral Research in Ocular Biometry and Emmetropisation. He co-owns a private independent practice in the Southern Highlands of NSW and is the Director of the Academy of Advanced Ophthalmic Optics, is a Fellow of the ABDO and ADOA, and was the 2022 International Optician of the Year. He is also the Past Chairman of the NSW Optical Dispensers Education Trust and Past Vice President of ADOA (NSW) and a current appointee to the Australian Standards Committee for Spectacles.

More reading

Assistance tools for patients with low vision

When spectacles aren’t enough

Fundamentals of ideal frame selection

Send this to a friend