Sub-Category Classifiers for Multiple-Instance Learning and its Application to Retinal Nerve Fiber Layer Visibility...

Area of Science:

• Ophthalmology
• Computer Vision
• Machine Learning

Background:

• Assessing retinal nerve fiber layer (RNFL) visibility in fundus images is crucial for diagnosing eye conditions.
• Current methods may lack precision in automated RNFL visibility assessment.

Purpose of the Study:

• To develop a novel multiple instance learning (MIL) method for automated RNFL visibility assessment in fundus camera images.
• To enable both image-level classification (visible/not visible) and region-level localization of RNFL in images.

Main Methods:

• A novel multiple instance learning (MIL) approach was developed, transforming the feature space to a discriminative subspace.
• A region-level classifier was learned within this subspace using a margin-based loss function for joint optimization.
• The method was trained and validated on a dataset of 884 fundus images annotated by ophthalmologists.

Main Results:

• The proposed MIL system achieved high agreement with ophthalmologist annotations (kappa values of 0.73 and 0.72).
• The system demonstrated better agreement with the more experienced annotator.
• Comparative tests on public datasets (MESSIDOR, DR, UCSB) showed superior performance over existing state-of-the-art methods.
• The developed Matlab code is publicly available for research use.

Conclusions:

• The novel MIL method effectively assesses RNFL visibility and localizes relevant regions in fundus images.
• The approach shows strong agreement with expert clinical judgment and outperforms current state-of-the-art techniques.
• This method holds promise for improving automated analysis of retinal images in clinical practice.