Classifying retinal images via vascular-optic disc cross-segmentation and attentive feature selection

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Retinal disorders are increasing, leading to blindness.
• Artificial intelligence (AI) offers potential for automated retinal image analysis.
• Challenges include image variability and imbalanced data, hindering robust model development.

Purpose of the Study:

• To develop a novel AI approach for classifying retinal images into healthy, glaucoma, and diabetic retinopathy categories.
• To improve classification performance by focusing on discriminative feature extraction and integrating multiview information.
• To address challenges like low inter-class variability and imbalanced datasets.

Main Methods:

• An abnormality-aware attentive feature selection method was employed.
• A cross-segmentation framework was utilized to extract and integrate optic disc and vascular structures.
• The approach was validated on multiple public datasets (FIVES, Drishti-GS1, SUSTech, HRF, PAPILA).

Main Results:

• The model achieved a balanced accuracy of 83% (95% CI: 80%-86%).
• High sensitivity, specificity, PPV, and NPV were reported for healthy, glaucoma, and diabetic retinopathy classifications.
• The method demonstrated robustness in distinguishing between the three classes across diverse datasets.

Conclusions:

• The proposed AI method effectively classifies retinal images for glaucoma and diabetic retinopathy detection.
• Integrating optic disc and vascular structures enhances classification accuracy and robustness.
• This approach shows promise for automated screening of critical retinal disorders.