Advances in machine learning for keratoconus diagnosis.

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Keratoconus (KC) diagnosis relies on various clinical and imaging data.
• Machine learning (ML) offers potential for improving KC detection and staging.
• A gap persists between academic ML research and its clinical application in ophthalmology.

Purpose of the Study:

• To review Machine Learning (ML) applications in keratoconus (KC) diagnosis over the last decade.
• To identify advancements, common methodologies, and challenges in ML for KC.
• To highlight the gap between ML research and clinical implementation.

Main Methods:

• Systematic literature search of digital libraries for ML in KC diagnosis.
• Inclusion/exclusion criteria applied to 62 identified articles.
• Analysis focused on ML algorithms, imaging modalities, datasets, and KC conditions studied.

Main Results:

• Supervised classifiers dominate KC diagnosis (97%), with Random Forest most frequent.
• Pentacam is the leading imaging modality (56%); most studies use local numerical data (91%).
• Research primarily compares normal vs. keratoconus eyes, with limited focus on KC severity stages (20%).

Conclusions:

• Lack of consensus on early KC detection standards and severity staging hinders progress.
• Limited multidisciplinary collaboration and restricted access to public datasets are key obstacles.
• Further research and roadmap models are needed for clinical integration of ML in KC diagnosis.