The Future of Keratoconus Screening with Artificial Intelligence
Hwang et al1 (page 1862) have combined corneal data from slit-scan tomography and spectral-domain OCT in a method for screening corneas for very early signs of keratoconus. The use of artificial intelligence (AI; or machine intelligence) has a decades-long history in corneal topography interpretation. The methods used have been discriminative classifiers that—given a set of independent machine-derived variables from corneal topography (e.g., simulated K readings, topographic asymmetries, etc.)—can be trained to distinguish between 2 or more classes of topography (e.g., normal, astigmatic, keratoconus). When using tomographic data rather than Placido topography, investigators have found it necessary to combine either biomechanical or additional imaging data to enhance the performance to detect early signs of keratoconus. Hwang et al used a direct statistical approach (multivariate logistic regression analysis) to tease out those variables available from slit-scan tomography and spectral-domain OCT that are statistically different between normal controls and the “clinically normal fellow eyes of highly asymmetric keratoconus subjects.” The best discrimination between the 2 groups was found when using a combination of variables from both instruments. The most significant parameters for discrimination arose from spectral-domain OCT corneal thickness measures followed by anterior corneal measures from tomography. Importantly, and contrary to widespread belief, the authors provided experimental evidence to show that “posterior corneal indices were not useful in distinguishing populations” (normal vs. asymmetric keratoconus).
The fine work by Hwang et al is a step forward toward understanding the importance of mixed data sets derived from more than one imaging method, and the authors intend to explore other methods such as the incorporation of biomechanical metrics as championed by Vinciguerra et al.2 Previous keratoconus detection systems were successful (100% sensitivity and specificity) using only Placido curvature data,3 and it is curious why that sensitive imaging technique is not used more often in screening studies, especially because the study by Hwang et al dismisses the importance of posterior corneal indices in detecting early signs of keratoconus.
There are significant hurdles in discriminating patients with very early signs of keratoconus from a normal population, not the least of which is the limited availability of samples owing to the low prevalence of the disease. As discussed by Hwang et al, before an objective test can be formulated, classification of pathologic features is accomplished subjectively by one or more experts. When experienced experts receive appropriate training, a consensus method can be used to produce validated data sets. Alternatively, the less affected cornea in a bilateral condition such as keratoconus can be used unambiguously to form a valid group, as done by Hwang et al. Nevertheless, a limited number of observations (30 in the Hwang et al study) may mean that the characteristics learned in AI training may not be similar to those in another clinic population, as pointed out by the authors. As a result, the performance of the discrimination can suffer from lack of exposure to a sufficient variety of presentation forms. For example, although keratoconus often presents as an inferior localized area of steepening, the location of steepening can occur in any corneal quadrant. Also, keratoconus can present as a pellucid marginal degeneration-like pattern (crab claw) with or without skewing of the radial axes, or as a foreshortened or truncated bow tie. It can be difficult to obtain examples of all these presentations as well as the many variations in between.
This challenge requires significantly more numbers of samples than generally are available using the same model of equipment to provide examples of the earliest manifestations of a disorder such as forme fruste keratoconus. Lack of a large sample size is a common problem with AI applications in medical diagnostics because many of the latest AI techniques are designed to mine big data sets that simply are not available for many low-incidence health conditions. Looking forward, special techniques in machine learning are being developed that generalize limited sample sets, expanding these into large numbers to provide the needed diversity. We look forward to the promise of such progress in machine learning to advance diagnostics not only in ophthalmology, but also in all fields of medicine.