Unsupervised learning for large-scale corneal topography clustering.

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Machine learning (ML) in medicine often requires extensive labeled data for supervised algorithms.
• Current ML applications primarily focus on mimicking human diagnostic capabilities.
• Vast unlabeled medical datasets, such as those from digital examinations, remain under-exploited.

Purpose of the Study:

• To demonstrate the efficacy of unsupervised learning in extracting and sorting usable data from large, unlabeled medical datasets.
• To apply unsupervised algorithms to corneal topography examinations for automated data categorization.
• To reduce human intervention in the data pre-processing pipeline for medical AI.

Main Methods:

• Application of unsupervised machine learning algorithms to a database of unlabeled corneal topography examinations.
• Automated sorting of extracted data into predefined diagnostic categories: Normal, Keratoconus, and History of Refractive Surgery.
• Utilizing existing digital examination tools and their stored databases.

Main Results:

• Successfully extracted 7019 usable corneal examinations from an unlabeled dataset.
• Achieved an overall accuracy of 96.5% in automatically sorting examinations into the three diagnostic categories.
• Demonstrated minimal human intervention required for data extraction and sorting.

Conclusions:

• Unsupervised learning offers a powerful method for pre-processing large unlabeled medical datasets, significantly speeding up data collection.
• This approach can facilitate the development of more robust supervised ML models by providing accurately sorted data.
• The methodology is adaptable to various digital examination databases, enhancing their utility.