Explainable machine learning framework for cataracts recognition using visual features.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Cataract is a primary cause of global blindness and visual impairment.
• Deep neural networks (DNNs) show promise in cataract recognition from AS-OCT images but lack clinical explainability.
• Visual features from AS-OCT images offer interpretability but remain underexplored.

Purpose of the Study:

• To develop an explainable machine learning framework for automated cataract severity recognition using AS-OCT images.
• To leverage interpretable visual features for improved clinical diagnostic practice.

Main Methods:

• Visual feature extraction from AS-OCT images and histograms using intensity-based statistical methods.
• Feature importance analysis and selection via SHapley Additive exPlanations and Pearson correlation coefficient.
• Ensemble multi-class ridge regression for cataract severity level recognition.

Main Results:

• The proposed framework achieves competitive performance compared to DNNs on the AS-OCT-NC dataset.
• The framework demonstrates strong explainability, aligning with clinical diagnostic needs.
• Selected visual features effectively contribute to accurate cataract severity classification.

Conclusions:

• The developed explainable machine learning framework offers a viable and interpretable approach for automated cataract grading.
• This method enhances the clinical applicability of AS-OCT imaging in ophthalmology.
• Combining visual features with explainable AI provides a robust solution for diagnosing cataract severity.