Focused Attention in Transformers for interpretable classification of retinal images

Area of Science:

• Ophthalmology
• Computer Vision
• Artificial Intelligence

Background:

• Vision Transformers (ViTs) are increasingly popular for image classification due to high performance and interpretability.
• These characteristics require thorough evaluation in the context of retinal imaging.
• Current attribution methods for ViTs produce low-resolution heatmaps, limiting their clinical utility.

Purpose of the Study:

• To compare the performance of various Vision Transformers against traditional Convolutional Neural Networks (CNNs) for retinal disease classification.
• To evaluate the models' ability to handle multi-modality retinal imaging (fundus and OCT) and generalize to external datasets.
• To introduce and validate a novel mechanism, Focused Attention, for generating high-resolution, interpretable attribution maps.

Main Methods:

• Performance evaluation of multiple Vision Transformer architectures versus CNNs on retinal disease classification tasks.
• Assessment of multi-modal imaging (fundus, OCT) and external data generalization.
• Development of Focused Attention, a novel attribution method using iterative conditional patch resampling.
• Validation of interpretability and lesion detection capabilities through a survey with four retinal specialists.

Main Results:

• Vision Transformers show competitive performance in retinal disease classification.
• Focused Attention generates higher-resolution attribution maps compared to existing Transformer methods.
• Retinal specialists found Vision Transformer attribution maps more interpretable than those from CNNs.
• Focused Attention was validated as a relevant tool for lesion detection in retinal images.

Conclusions:

• Vision Transformers offer a promising alternative to CNNs for retinal disease classification, with enhanced interpretability.
• The proposed Focused Attention mechanism significantly improves the resolution and clinical relevance of attribution maps.
• This work highlights the potential of Vision Transformers and advanced interpretability techniques in ophthalmology.