Structure-Oriented Transformer for retinal diseases grading from OCT images.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Retinal diseases are a primary cause of vision loss, necessitating accurate grading for timely intervention.
• Current Convolutional Neural Network (CNN) methods struggle with long-range dependencies crucial for understanding lesion-retina relationships.

Purpose of the Study:

• To propose a novel Structure-Oriented Transformer (SoT) framework for improved retinal disease grading.
• To enhance the modeling of relationships between local retinal lesions and the overall retinal structure.

Main Methods:

• Developed a Structure-Oriented Transformer (SoT) incorporating structure guidance for emphasizing global retinal context.
• Utilized a pre-trained vision transformer for efficient modeling of feature patch relationships via transfer learning.
• Implemented a Token vote classifier to aggregate information from all output tokens for final grading.

Main Results:

• The SoT framework demonstrated effectiveness in constructing lesion-retina relationships on clinical datasets.
• Experiments on a neovascular Age-related Macular Degeneration (nAMD) dataset showed superior performance compared to state-of-the-art methods.
• Evaluation on a public retinal diseases dataset confirmed the algorithm's classification superiority and generalizability.

Conclusions:

• The Structure-Oriented Transformer (SoT) significantly improves the ability to relate local lesions to the whole retina for accurate grading.
• SoT offers a robust and generalizable approach for retinal disease classification, outperforming existing techniques.