Automatic detection of referral patients due to retinal pathologies through data mining

Area of Science:

• Ophthalmology
• Medical Imaging
• Data Mining

Background:

• Increasing prevalence of retinal pathologies necessitates advanced diagnostic tools.
• Current automated detection algorithms often focus on single pathologies like diabetic retinopathy, limiting clinical utility.
• Clinicians require tools that can assess overall retinal health, considering multiple pathologies simultaneously.

Purpose of the Study:

• To develop a comprehensive algorithm for characterizing both normal and abnormal retinal appearances from examination records.
• To address the limitations of single-pathology detection systems by incorporating contextual patient information and multiple retinal images.
• To create a flexible and adaptable system for automated diagnosis of various retinal conditions.

Main Methods:

• A novel data mining approach is employed to learn diagnosis rules from fundus examination records.
• The algorithm characterizes retinal data at multiple levels of spatial and lexical granularity.
• Adaptive decomposition of retinal images into regions and feature spaces into visual words ensures flexibility.

Main Results:

• The framework was evaluated on two large datasets (e-ophtha and Messidor), comprising over 25,000 examination records.
• The algorithm successfully identified patients requiring referral to an ophthalmologist.
• It also demonstrated effectiveness in the specific detection of multiple retinal pathologies.

Conclusions:

• The proposed multigranular approach offers a flexible method for characterizing retinal normality and abnormality.
• The data mining variation enables effective analysis of contextual and visual data at adaptive granularities.
• This comprehensive system enhances automated eye screening by detecting diverse pathologies and identifying patients needing specialist care.