Deep Ensemble Learning for Retinal Image Classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Vision impairment affects 2.2 billion globally, with many cases preventable through early detection.
• Current automated screening using CNNs is limited to a few ocular pathologies.
• Simultaneous detection of multiple pathologies can enhance clinical utility.

Purpose of the Study:

• To develop and evaluate an ensemble of CNNs for simultaneous detection and classification of multiple ocular pathologies from retinal fundus images.
• To improve the accuracy and clinical usability of automated ophthalmic pathology screening.

Main Methods:

• Utilized the Retinal Fundus Multi-Disease Image Dataset (RFMiD) with 2,560 images.
• Trained five CNN architectures to predict pathology presence and classify 28 distinct pathologies.
• Employed an ensemble approach, averaging predictions from individual models, trained to minimize asymmetric loss.

Main Results:

• The ensemble network achieved a high AUROC of 0.9613 for disease screening (healthy vs. pathologic).
• Individual disease classification yielded an average AUROC of 0.9295 across all classes.
• The ensemble model outperformed the best single CNN architecture (SE-ResNeXt, AUROC 0.9586).

Conclusions:

• Ensemble CNNs effectively detect and classify multiple ocular pathologies from retinal images, surpassing individual model performance.
• This deep learning strategy enhances automatic screening and diagnosis of ophthalmic conditions.
• External validation is crucial for clinical translation of these machine learning models.