Role of Artificial Intelligence in Diagnosis and Management of Keratoconus

Sanjukta Jana, B.Optom

Intern-Optometrist, LVPEI – Bhubaneswar, India

 

Across all health paradigms, early disease diagnosis and its management will outsmart the long-time health hazards for our patients. Nowadays, machine learning techniques have been introduced for ophthalmic screening and in detailed analysis of a range of conditions including keratoconus.

A question may arise in readers’ mind that – Why early detection of keratoconus is important?

Early identification of both at-risk patients and those likely to progress is vital to maintain best-corrected visual acuity.

Machine Learning and Keratoconus

Machine learning techniques have been utilised across ophthalmic screening and analysis for a range of conditions including glaucoma, diabetic retinopathy, neovascular macular degeneration, uveitis, cataract surgery, and visual field interpretation.

To fully appreciate the potential role of machine learning within the keratoconus paradigm, it is necessary to have a basic understanding of artificial intelligence (Figure 1).

Figure 1: Relationship between artificial intelligence, machine learning, and deep learning.
_{[Picture courtesy: https://ai.stackexchange.com/questions/15859/is-machine-learning-required-for-deep-learning ]}

Introduction to Artificial Intelligence and its two components: Machine learning and Deep learning

AI emerged as a simple theory as early as the 1950s, described as “giving computers the ability to learn without being programmed.” In medical terms, the appropriate definition of AI would be “a system’s ability to interpret data and to properly utilise the findings to achieve specific tasks through adaptation.”

Artificial intelligence represents the overarching term, with machine learning and then deep learning being subsets within this broad area.

Machine learning (ML) is the utilisation of specific algorithms that are trained to adapt to patterns in data by exposure and experience retained over time. Artificial intelligence represents the overarching term, with machine learning and then deep learning being subsets within this broad area. (Figure 2) ⁽¹⁾

Machine learning can be supervised, semi-supervised, or unsupervised:

• Supervised learning identifies patterns to relate variables to measured outcomes. Supervised learning, as the name suggests, is led by the researcher who provides the variables and/or research question.
• Semi-supervised learning utilises labels provided by the researcher to infer classification or regression through the subsequent analysis.
• Unsupervised learning generally seeks to find patterns in unlabelled data and is external to researcher input. ⁽²⁾

Figure 2: Depicting comparison between machine learning and deep learning.
_{[Picture Courtesy: https://levity.ai/blog/difference-machine-learning-deep-learning]}

Limitations of Artificial Intelligence Programs

The success of AI programs will be limited by access to information and the understanding of the condition investigated.

Keratoconus and Artificial Intelligence

The focus of most current AI applications has been the diagnosis and classification of the condition. In the diagnosis of keratoconus, various AI models have been used which include – neural networks, support vector machines, and decision trees. ⁽³⁾ In artificial neural networks, the main aim is to extract data and compute them into successive networks. ⁽³⁾

In decision trees, data is divided into various structures depending on the set of questions. ⁽³⁾ In a support vector machine there is maximum possible separation of data through the optimal boundary. To figure out the optimal boundary the data are mapped into a 3D configuration. ⁽³⁾ Corneal topography and tomography are the gold standard evaluation for keratoconus diagnosis. ⁽⁴⁾ In 1995, Maeda, Klyce, and Smolek described preliminary findings of a neural network trained on 11 parameters from the TMS-1 video keratoscope that characterised corneal shape. ⁽⁴⁾

Stromal thinning is a feature of keratoconus. The Ambrosio relational thickness measure represented one of the first models adding corneal thickness to algorithms and has been shown to provide high specificity (85.5 to 100%) and sensitivity (88.9 to 99%) across multiple studies highlighting the benefits of incorporating pachymetry.⁽⁵⁾

Conclusion:

Variable corneal properties continue to impact the predictability of treatment outcomes in keratoconus patients. Utilising AI-based models may provide specialists with a more optimal approach to interventions including intrastromal corneal ring segments.

 

References

1. Gokul A, Patel DV, McGhee CN. Dr John Nottingham’s 1854 landmark treatise on conical cornea considered in the context of the current knowledge of keratoconus. Cornea.
2. Mooney SJ, Pejaver V. Big data in public health: terminology, machine learning, and privacy. Annu Rev Public Health. 2018
3. Artificial intelligence in keratoconus- https://entokey.com/artificial-intelligence-in-keratoconus/ Role of Artificial Intelligence in diagnosis and management of Keratoconus by Sanjukta Jana
4. Martínez-Abad A, Piñero DP. New perspectives on the detection and progression of keratoconus. J Cataract Refract Surg. 2017
5. Kobashi H, Tsubota K. Accelerated versus standard corneal cross-linking for progressive keratoconus: a meta-analysis of randomised controlled trials.Cornea. 2020