Explainable ensemble learning method for OCT detection with transfer learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) in optical coherence tomography (OCT) image detection is vital for reducing clinician workload and improving diagnostic accuracy.
• Explainability and accuracy are key for advancing AI in clinical workflows, particularly in retinal imaging.

Purpose of the Study:

• To develop and evaluate an explainable ensemble AI approach for detecting fundus diseases in OCT images.
• To assess the impact of transfer learning and pre-trained weights on AI model performance for OCT-based disease detection.

Main Methods:

• Utilized a public OCT dataset with normal subjects, dry age-related macular degeneration (AMD), and diabetic macular edema (DME) samples.
• Employed transfer learning with pre-trained ImageNet weights, comparing individual network performance before ensembling via majority soft polling.
• Visualized learned features using Grad-CAM and CAM for interpretability.

Main Results:

• Pre-trained ImageNet weights significantly improved individual network performance from 68.17% to 92.89%.
• The ensemble model achieved 100% accuracy in distinguishing between AMD, DME, and normal subjects.
• Grad-CAM visualization demonstrated accurate identification of lesion areas.

Conclusions:

• The proposed explainable ensemble AI approach demonstrates high accuracy and interpretability for retinal OCT image detection.
• Transfer learning and ensemble methods are effective strategies for enhancing AI performance in diagnosing retinal diseases.
• This approach shows potential for streamlining clinical workflows in ophthalmology.