Ameerah Ilyas
Introduction
Artificial intelligence (AI) has redefined the diagnosis and early detection of multiple ocular conditions in the realm of ophthalmology. Artificial intelligence was initially focused on retinal conditions, such as the grading and early detection of diabetic retinopathy and age related macular degeneration through the use of optical coherence tomography (OCT) and retinal databases (1). Systems such as IDx-DR achieved FDA approval for the detection of diabetic retinopathy, and based on a set of protocols was able to determine those that required further follow up (2). This paved the way for further systems being introduced for screening and now for anterior segment conditions such as keratoconus, becoming the focal point of its expansion(1,3).
Keratoconus
Keratoconus is a bilateral, asymmetrical corneal ectatic disorder characterized by progressive corneal steepening and stromal thinning (4) and occurs within the second and third decades of life, these pathological corneal changes can lead to a reduction in visual acuity and irregular astigmatism if not detected early with intervention (5). Hence the importance of early detection within the subclinical stages, to ensure cross linking can be done to prevent its progression (6).
It is also important for prospective patients who are undergoing refractive surgery, as it can potentially trigger iatrogenic ectasia (7). Corneal topography remains the most effective method of disease detection as well as progression (8). Deep learning and applied machine learning models have been used to detect the early signs of keratoconus on topography and OCT images (9) with algorithms ranging from decision trees and support vector machines showing diagnostic accuracy and specificity of 0.95, with a few studies nearing 1.00 (1).
Future of keratoconus detection
A systematic review demonstrated that machine learning that can incorporate a variety of corneal parameters and biometric measures such as visual acuity and refraction displayed >90% specificity and >80% sensitivity. Studies which included corneal biomechanical measurements and epithelial thickness grading were more accurate (9). With more advanced systems involving data such as age, family history, ethnicity into the predictive input models, demonstrating the multifactorial aspect of the disease (7,10).
Limitations of AI
While AI has shown promising results in terms of early detection, there are important considerations before implementing on a larger scale and in regular practice. Most studies have incorporated data from single centre data sets and hence may not include outliers or with the whole spectrum of the disease severity, as well as ethnicity bias depending on the region (10,11).
Furthermore, defining subclinical keratoconus has been widely disputed, with no formal set parameters to define such cases despite attempts such as Belin ABCD grading or Amsler-Krumeich classification. This makes creating algorithms for thresholds of parameters difficult and when further follow up or intervention is required. This ultimately means even systems showing high levels of accuracy may be implementing biases into the training label (10).
Another hurdle in its application to daily practice involves output data that is easily interpreted by clinicians and complies with the trust regulatory and data privacy regulations (5,8). More so, in deep learning models, the system highlights potential subclinical cases with no determining factor behind the result, which can lead to mistrust by the clinician. On the other hand, becoming over reliant on the models could potentially reduce diagnostic vigilance as compared to standard care due to the multifactorial nature of the disease (5,9,11). Particularly in patients with advanced keratoconus and other ocular conditions such as corneal scarring which may alter corneal parameters and lead to misleading results.
Conclusions
Overall AI is showing a promising future in detecting subclinical cases of keratoconus often overlooked by human interpretation. With more advanced models including a wider range of input data, the accuracy and specificity of models is becoming increasingly more precise. The challenge will arise when shifting from clinical research and including a wider range of datasets with set corneal parameters and external regulation into clinical practice (6,9,10).
The results of AI systems being implemented in the screening of diabetic retinopathy, drusen quantification in age related macular degeneration as well retinopathy of prematurity (1,2) shows the potential in not only keratoconus detection but also other corneal manifestations such as infectious keratitis and pterygium (3). The objective of these systems is not to replace clinicians but to speed up the process of detection and early involvement of specialists, particularly in low income and high burden countries (5).
References
1. Ji Y, Liu S, Hong X, et al. Advances in artificial intelligence applications for ocular surface diseases diagnosis. Front Cell Dev Biol. 2022;10:1107689. doi:10.3389/fcell.2022.1107689
2. Hashemian H, Peto T, Ambrosio R, et al. Application of artificial intelligence in ophthalmology: an updated comprehensive review. J Ophthalmic Vis Res. 2024;19(3):15893. doi:10.18502/jovr.v19i3.15893
3. Zhang Z, Wang Y, Zhang H, et al. Artificial intelligence-assisted diagnosis of ocular surface diseases. Front Cell Dev Biol. 2023;11:1133680. doi:10.3389/fcell.2023.1133680
4. Cunha, A.M., Correia, P.J., Alves, H. et al. Keratoconus enlargement as a predictor of keratoconus progression. Sci Rep 11, 21079 (2021). https://doi.org/10.1038/s41598-021-00649-0
5. Li X, Rabinowitz YS, Rasheed K, Yang H. Longitudinal study of the normal eyes in unilateral keratoconus patients. Ophthalmology. 2004 Mar;111(3):440-6. doi: 10.1016/j.ophtha.2003.06.020. PMID: 15019316.
6. Deshmukh R, Ong ZZ, Rampat R, et al. Management of keratoconus: an updated review. Front Med. 2023;10:1212314. doi:10.3389/fmed.2023.1212314
7. Ambrosio R, Salomao MQ, Barros L, et al. Multimodal diagnostics for keratoconus and ectatic corneal diseases: a paradigm shift. Eye Vis. 2023;10:45. doi:10.1186/s40662-023-00363-0
8. Hashemi H, Mehravaran S. Day to Day Clinically Relevant Corneal Elevation, Thickness, and Curvature Parameters Using the Orbscan II Scanning Slit Topographer and the Pentacam Scheimpflug Imaging Device. Middle East Afr J Ophthalmol. 2010 Jan;17(1):44-55. doi: 10.4103/0974-9233.61216. PMID: 20543936; PMCID: PMC2880373.
9. Vandevenne MMS, Favuzza E, Veta M, et al. Artificial intelligence for detecting keratoconus. Cochrane Database of Systematic Reviews. 2023;(11):CD014911. doi:10.1002/14651858.cd014911.pub2
10. Maile H, Li JPO, Gore DM, et al. Machine learning algorithms to detect subclinical keratoconus: systematic review. JMIR Med Inform. 2021;9(12):e27363.doi:10.2196/27363
11. Tey KY, Cheong EZK, Ang M. Potential applications of artificial intelligence in image analysis in cornea diseases: a review. Eye Vis. 2024;11:10. doi:10.1186/s40662-024-00376-3