An explainable transformer model for Alzheimer's disease detection using retinal imaging.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Ophthalmology

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder affecting millions globally.
• Early diagnosis of AD is critical for timely management to delay disease progression.
• Current diagnostic methods have limitations, necessitating novel approaches.

Purpose of the Study:

• To introduce Retformer, a novel transformer-based AI architecture for early Alzheimer's disease detection using retinal imaging.
• To leverage explainable AI (XAI) to understand the model's diagnostic reasoning.
• To compare Retformer's performance against established algorithms.

Main Methods:

• Retformer was trained on diverse retinal image modalities from AD patients and healthy controls.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used for visualising feature importance.
• Model performance was evaluated against benchmark algorithms using various metrics.

Main Results:

• Retformer effectively learned complex patterns in retinal images for AD diagnosis.
• Explainable AI highlighted key retinal regions crucial for AD detection.
• The model demonstrated superior performance, outperforming benchmarks by up to 11%.

Conclusions:

• Retformer shows significant promise as an accurate and explainable tool for early Alzheimer's disease detection via retinal imaging.
• Retinal imaging combined with advanced AI offers a viable strategy for non-invasive AD diagnosis.
• Further research can refine Retformer for clinical application in AD management.